Skip to main content

icelake

Persistent memory for Discord bots. It passively reads messages, extracts facts about your members, and hands you a labeled block to paste into your system prompt when the bot needs to reply.

Facts are stored against Discord user IDs, not names, so a rename never moves someone's memories onto the wrong person. Third-party claims ("alice called bob a hacker") attach to the person they are about.

Requires Python 3.12+.

Install

pip install icelake
pip install "icelake[discord]"           # discord.py helpers/hooks
pip install "icelake[mongo]"             # MongoDB backend
pip install "icelake[local-embeddings]"  # sentence-transformers

Quickstart

import asyncio
from datetime import UTC, datetime

from icelake import Audience, DiscordMemory, MemoryConfig, MessageEvent


async def main() -> None:
    memory = DiscordMemory(
        MemoryConfig(
            storage="sqlite:///memory.db",
            llm="openai://$OPENROUTER_API_KEY@openrouter.ai/api/v1?model=z-ai/glm-5.3-flash&reasoning=low",
        )
    )
    async with memory:
        await memory.observe(
            MessageEvent(
                message_id="9001",
                guild_id="555",
                channel_id="777",
                author_id="100000000000000001",
                content="I've been learning Rust for about a year now!",
                created_at=datetime.now(UTC),
                author_display_name="alice",
            )
        )

        ctx = await memory.prompt_context(
            guild_id="555",
            audience=Audience(channel_id="777", requester_id="100000000000000001"),
            asker_id="100000000000000001",
            text="what am I learning these days?",
            mentioned_ids=("200000000000000002",),
            thread_participant_ids=(),  # other humans in the reply chain / channel tail
        )
        print(ctx.injection_block)

        # Exact restatements can be cited without another provider call.
        # Paraphrases remain uncited unless you supply a CitationVerifier.
        reply = "You're learning Rust!"
        cited = await memory.cite(reply, ctx)
        reply = cited.text



asyncio.run(main())

observe returns immediately. Extraction runs in the background. prompt_context builds a labeled block for the asker, anyone they mentioned, thread participants, and the server. Mentions plus thread participants turn on graph-hop recall and pair-intersect every combination of those people (not just asker-other), so shared entities surface in one call. Stick the block on your system prompt and generate a reply. The model sees plain facts, never citation syntax. Then memory.cite(reply, ctx) links exact restatements to registered evidence. Uncertain claims stay uncited. Pass a verifier when you want citations for paraphrases.

Citations: provenance with explicit support checks

Citations registers the closed set of sources a reply may cite. Code owns URLs and source identities; the verifier can only select candidate IDs. memory.cite(reply, ctx) defaults to exact restatements, with no provider calls. Similarity alone cannot establish support: changing a person, negation, time, or quantity can leave an embedding almost unchanged.

For paraphrases, inject a CitationVerifier. The supplied LlmCitationVerifier uses your existing ChatLLM interface, a validated output schema, a 15-second deadline, and at most one schema-repair call. It checks at most 12 claims and three candidate sources per claim. retrieval.citation_min_score controls candidate selection, never proof of support. Supply a metered LLM to account for calls under the citation purpose.

from icelake import Citations, LlmCitationVerifier

citations = Citations.from_prompt_context(ctx)
citations.add_source(
    "https://example.com/docs", title="Docs", excerpt="The source passage goes here."
)
# verifier_llm is your implementation of the ChatLLM port.
verifier = LlmCitationVerifier(verifier_llm, guild_id=guild_id)
cited = await memory.cite(model_output, citations, verifier=verifier)

A citation establishes provenance, not independent truth. Model-based verification is fallible; missing, invalid, or failed verification leaves uncertain claims uncited. Third-party reports must retain their attribution. Provider-supplied character anchors can also be registered explicitly; they are trusted provenance annotations and do not pass through the claim verifier.

Rendering validates registered sources and verbatim span offsets before inserting [[N]](<url>). AttributedReply.claims and .sources contain used-only evidence. Pass weave=False for your own renderer. Footer numbers must use Citations.citable to agree with inline labels. See examples/citations.py.

Examples:

The current recommended extractor is z-ai/glm-5.3-flash. See Model benchmark for the ranked table and how to add a new row.

Model benchmark

One full run of examples/e2e_simulation.py per model (suite A drain-mode + suite B worker-mode) against OpenRouter on 2026-08-29 and 2026-08-30. Hard checks are library guarantees. Expectations are model-decided extraction/reconcile/classify outcomes. Spend is the meter's provider-reported USD. List prices are OpenRouter prompt / completion per 1M tokens as of 2026-08-30.

n=2 rows average the latest aug30-mapped run with the previous published fair run on comparable knobs. n=1 is either a first fair run (correct MODELS knobs in examples/bench_models.py) or a holdover not in this matrix. Not averaged: aug30-temp-none GLM (wrong knobs), GPT-5 mini without reasoning=minimal, Luna with reasoning=low.

Pick z-ai/glm-5.3-flash unless you have a reason not to. It hit 47/47 this run and still sits around a cent. Gemini 3.7 Flash also hit 47/47, at ~8x the spend. Mistral Small 3.2 is the cheap quality sleeper (43/47, retired both contradictions, $0.0064) but takes ~9.5 min. GPT-5 mini is usable with reasoning=minimal + temperature none, but still leaves Omaha live. 4o-mini ranks high on cost/speed while failing reconcile.

Rank Model Score n Exp. Hard Spend Time In $/M Out $/M
1 z-ai/glm-5.3-flash 94.0 2 46.5/47 83/83 $0.0079 3:56 $0.075 $0.25
2 openai/gpt-5.6-luna 89.0 1 40/46 83/83 $0.0121 1:17 $0.20 $1.20
3 mistralai/mistral-small-3.2-24b-instruct 88.3 1 43/47 83/83 $0.0064 9:32 $0.075 $0.20
4 openai/gpt-5-mini 85.4 1 41/46 83/83 $0.0236 1:54 $0.25 $2.00
5 openai/gpt-4o-mini 84.9 2 33/46 83/83 $0.0065 1:02 $0.15 $0.60
6 google/gemini-3.7-flash 82.8 2 45/47 83/83 $0.0672 2:06 $0.75 $3.75
7 openai/gpt-4.1-mini 82.1 2 36/47 83/83 $0.0185 1:13 $0.40 $1.60
8 openai/gpt-4.1-nano 82.0 1 30/46 83/83 $0.0056 0:56 $0.10 $0.40
9 openai/gpt-oss-120b 80.6 1 33/46 83/83 $0.0061 4:38 $0.037 $0.17
10 openai/gpt-5-nano 79.8 1 34/46 82/83 $0.0032 1:06 $0.05 $0.40
11 qwen/qwen3-30b-a3b-instruct-2507 79.2 1 32/46 83/83 $0.0054 5:12 $0.048 $0.193
12 inception/mercury-2 78.3 1 32/46 83/83 $0.0234 0:46 $0.25 $0.75
13 qwen/qwen3-32b 78.1 2 34/46 83/83 $0.0074 14:05 $0.08 $0.28
14 deepseek/deepseek-v4-flash 78.0 2 36/47 83/83 $0.0092 14:47 $0.081 $0.16
15 google/gemini-3.1-flash-lite 77.7 2 42/47 82/83 $0.0593 2:11 $0.25 $1.50
16 google/gemini-2.5-flash 76.8 2 37/47 83/83 $0.0471 1:56 $0.30 $2.50
17 minimax/minimax-m3 75.9 1 36/46 82/83 $0.0184 1:32 $0.30 $1.20
18 google/gemini-3-flash-preview 75.4 1 43/47 82/83 $0.0612 1:59 $0.50 $3.00
19 deepseek/deepseek-v4-flash-0731 72.7 2 30/46 83/83 $0.0111 13:05 $0.065 $0.18
20 x-ai/grok-4.3 66.5 1 33/46 83/83 $0.1012 4:20 $1.25 $2.50
21 ~anthropic/claude-haiku-latest 66.4 1 37/46 83/83 $0.2160 5:08 $1.00 $5.00
22 google/gemma-4-31b-it 64.9 1 28/45 83/83 $0.0224 39:42 $0.09 $0.34
23 z-ai/glm-4.5-air 57.1 1 24/45 83/83 $0.0624 11:15 $0.13 $0.85
24 qwen/qwen3.7-flash 55.9 1 17/45 82/83 $0.0092 5:07 $0.03 $0.13
25 tencent/hy4-preview 23.3 1 8/44 82/83 $0.2871 23:18 $0.83 $2.50

Score is /100. In/out are OpenRouter list prices, not the run. Expectation denominators differ slightly when a model never produced the fixture a later check needs. Spend under the $0.007 cost floor (4o-mini, 4.1-nano, gpt-oss-120b, gpt-5-nano, Qwen3-30B) clips the cost component to 100.

Score

Fixed anchors, so a new row does not rescale the others:

Exp   = 100 * (expectations met) / (expectations total)          # suites A+B
Hard  = 100 if zero hard failures, else max(0, 100 - 30 * failures)
        # n>1: use mean failure count
Cost  = 100 * (1 - log10(spend / 0.007) / log10(0.30 / 0.007))   # clamp 0-100
Speed = 100 * (1 - log10(seconds / 45)  / log10(1400 / 45))      # clamp 0-100

Score = 0.50*Exp + 0.20*Hard + 0.20*Cost + 0.10*Speed

$0.007 / $0.30 and 45s / 1400s are this round's observed best/worst. Leave them unless a new run is clearly outside the band.

To add a model: run uv run python examples/bench_models.py --models <id> --out bench_runs/<date>. If it already has a fair run on the same knobs, average hard / expectations / spend / duration. Otherwise use the new run alone. Plug current OpenRouter in/out, compute Score, insert the row in rank order.

What actually differed

The discriminating job is still reconcile, not first-pass extraction. Almost every model stored "lives in Omaha" and "loves Red Bull". This mapped run, only GLM 5.3 Flash, Gemini 3.7 Flash, Mistral Small 3.2, Gemini 3 Flash Preview, and Gemini 3.1 Flash Lite retired both when Alice moved to Seattle and quit the drink. GPT-5 mini and Luna retired Red Bull but left Omaha live. Gemini 2.5 Flash and GPT-4.1-mini did it on an earlier run and did not this time.

Name-binding is the other split: Bob saying "nolan's last name is gregory" must not become Bob's alias. GPT-5 nano, Gemini 3.1 Flash Lite, Gemini 3 Flash Preview, and MiniMax M3 failed that hard check.

Model What it did well What it missed
GLM 5.3 Flash Perfect 47/47 this run (retired Omaha + Red Bull). n=2 still ~a cent First run missed the charge-nurse merge
GPT-5.6 Luna First fair with reasoning=none: 40/46, $0.012, 77s. Retired Red Bull Omaha still live. Nursing versions piled up. Missed piano age-flush
Mistral Small 3.2 First fair. Retired both contradictions at $0.0064 ~9.5 min. Leftover "prefers yellow cans" next to "does not drink". No puppy merge / charge nurse / drums. Two reconcile JSON drops in the log
GPT-5 mini First fair with minimal + temp none. 41/46, $0.024, 1:54. Retired Red Bull. Opening batch survived Omaha still live. Nursing versions piled up. Do not average the earlier $0.11 disaster (no minimal. Dropped Alice's first batch)
GPT-4o-mini n=2 cheapest-completed band, ~1 min. Mem0/Cognee's old default Omaha and Red Bull still live both runs. No puppy / purple
Gemini 3.7 Flash Perfect 47/47 this run (also retired both). Fast (~2 min) n=2 spend ~8x GLM. First run missed a classify query
GPT-4.1-mini Graphiti's default. Fast Inconsistent: first run retired Omaha, this run left Omaha + Red Bull live
GPT-4.1-nano First fair. Fastest OpenAI row (56s), $0.0056 Omaha and Red Bull still live. Thin store. No puppy / purple / piano
gpt-oss-120b First fair. GLM-level spend 4:38. Omaha and Red Bull still live. One extraction JSON drop
GPT-5 nano First fair (json_object no longer needed). Cheapest run ($0.0032) Hard fail: Bob stating Nolan's surname became Bob's alias (gregory). Never extracted Red Bull. Added "the move went well" without retiring Omaha
Qwen3-30B instruct First fair. $0.0054 5:12. Omaha and Red Bull still live. Two extraction JSON drops
Mercury 2 Fastest by far (46s). Holdover Same reconcile misses as the pack
Qwen3 32B GLM-level spend both runs ~14 min. Omaha and Red Bull still live
DeepSeek V4 Flash This run retired Omaha (not Red Bull). Near-GLM spend n=2 time pulled to ~15 min by the 1189s mapped run. Two extraction JSON drops. 0731 still worse
Gemini 3.1 Flash Lite This run retired both contradictions. 44/47 expectations Hard fail this run (same Gregory-alias miss). n=2 Hard 82/83. First run did not reconcile Omaha
Gemini 2.5 Flash First run retired both This run left Omaha live (Red Bull retired). n=2 quality dropped
MiniMax M3 Fast and cheap. Holdover Hard fail: Gregory alias on Bob. Omaha and Red Bull still live
Gemini 3 Flash Preview First fair. Retired both. 43/47 Hard fail: Gregory alias on Bob. Classify JSON dropped (remember / query missed)
DeepSeek 0731 Cleaner than its first published run (33/46 this time) Still no reconcile. Several JSON drops. ~13 min
Grok 4.3 Holdover Thin store. Expensive. Reconcile misses
Claude Haiku Latest Holdover. Same expectation rate as old Luna 28x GLM's spend. 401s and invalid JSON in the log
Gemma 4 31B Holdover. Clean hard pass 40 minutes. Missed name/Go/Omaha on first pass
GLM 4.5 Air First fair (json_object, was 404) Dropped Alice's opening batches (invalid JSON after one repair): no name / Go / Omaha / Red Bull / nursing. Later quit-Red-Bull facts attached to the wrong speaker
Qwen 3.7 flash First fair (json_object, was 404) Same opening-batch drops. Hard fail: BatchCompleted never fired. 6 facts
Hy4 preview Holdover Structured output kept failing. 5 rows, all curation probes. Failed BatchCompleted

Did not finish

Model In $/M Out $/M Why
openai/gpt-5.6-sol $2.00 $10.00 HTTP 404: no endpoint accepted the requested knobs
tencent/hy-mt2-1.8b $0.044 $0.18 Same 404 (translation model)
tencent/hy-mt2-30b-a3b $0.074 $0.30 Same 404 (translation model)

Mapped knobs live in MODELS in examples/bench_models.py: mandatory-thinking models get the cheapest allowed effort (low / minimal), GPT-5.x omits temperature, models that 404 on json_schema use structured_outputs=json_object. Holdovers are aug29-full-models. Challengers averaged into n=2 are aug30-challengers.

Graph explorer

A self-contained HTML canvas over the public API: users, entities, typed relations, identity links (entity is this member), and server facts. Incidence links (dm_links) are an index, not a relationship, and are not drawn.

python -m icelake.visualizer \
  --storage "mongodb://127.0.0.1:27017/icelake" \
  --guild YOUR_GUILD_ID \
  --out icelake-graph.html \
  --serve --open

--center klim --depth 2 limits the canvas to that neighborhood (server / the server works). --list-guilds prints guilds in storage. Serve over HTTP (--serve) so the layout worker can run. Opening the file directly still works with a one-shot layout.

A reply turn

ctx = await memory.prompt_context(
    guild_id=guild_id,
    asker_id=str(message.author.id),
    text=question,
    mentioned_ids=("alice_id", "bob_id"),
    thread_participant_ids=("carol_id",),
)

# ctx.injection_block looks like:
#
#   [MEMORY CONTEXT]
#
#   WHAT I KNOW ABOUT THE CURRENT ASKER
#   Facts about the asker ONLY:
#   - alice mains support in every ranked game she plays
#
#   REFERENCED USER: bob
#   Facts about bob ONLY. Do NOT attribute these to the asker.
#   - bob was called a hacker by alice during the ranked match
#
#   SERVER COMMUNITY FACTS
#   Community-wide traits:
#   - the community bonds over late night gaming sessions
#
#   (usage guidelines follow - no citation syntax anywhere)

reply = await generate(system_prompt + "\n\n" + ctx.injection_block, question)
cited = await memory.cite(reply, ctx)
await message.reply(cited.text, mention_author=False)
# cite weaves [[N]](<url>) at attributed claim spans. Unsupported claims
# stay uncited, and sources the reply never used never render.

prompt_context pair-intersects every combination of the asker, mentions, and thread participants (not just asker-other) and turns on graph-hop recall whenever anyone else is in the turn. The pairing helper is discovery_pairs(asker_id, mentioned_ids, thread_participant_ids) (exported from icelake) if you need the same pairs without a recall.

Identity is always memory.identity.resolve(guild, name_or_mention_or_id). Ambiguous matches never auto-pick. identity.display_name is the reverse map so replies can attribute by the name the guild actually uses.

"What do you know about X?" - names in prose

prompt_context is mention-and-thread keyed: it scopes sections from asker_id, mentioned_ids, and thread_participant_ids. It never scans the question text for names. Recall makes no LLM calls, by design. So when someone asks "what do you know about klim?" and klim is typed rather than @mentioned (and isn't in the thread), no klim section exists unless you resolve the name yourself.

Resolve against current Discord membership, then retrieve through the audience boundary. The live example asks for a mention when a name is ambiguous. For a strict profile, retain only facts whose subject is the resolved member:

result = await memory.recall(RecallQuery(
    guild_id=guild_id,
    subject_ids=(resolved_member_id,),
    audience=Audience(channel_id=reply_channel_id, requester_id=requester_id),
))
profile = [item.fact for item in result.facts
           if item.fact.subject_id == resolved_member_id]

Getting name out of the conversation is your side, and no tool runtime is required. Use whatever your bot already has:

  • A slash command with a free-text option (zero LLM). /memory lookup klim hands you the name directly. See memory_lookup in examples/omni_style_bot.py.
  • One structured-output call. A tiny JSON router via ChatRequest.response_schema classifies the turn and extracts the name, no function-calling machinery. See _route_name_lookup in the same file.
  • Native function calling. If your LLM client already speaks tools, hand it a schema and dispatch to the same handler.

Two details matter for accuracy:

  • Never guess on ambiguity. Resolution.ambiguous means more than one member matches. Ask which one. The ladder ranks mention ID > username > real name > display name > nicknames, and only auto-resolves a clear winner.
  • Fetch strict when answering "about X". Recall channels deliberately surface facts that touch a person, not just facts they are the subject of: a claim like "carol and bob argued over game night" is stored on carol, but bob sits on its incidence links (as actor or @mention), so recall(subject_ids=(bob,)) returns it under bob. That is right for passive context (the injection block labels who each fact is about), but for a direct profile answer, filter the authorized results by subject_id. facts.list_for_subject is available to trusted tooling, but does not enforce channel permissions.

Trusted offline identity/profile inspection: examples/name_lookup_tool.py. Wired into a live reply loop and a slash command: examples/omni_style_bot.py.

Names, relationships, commands

resolution = await memory.identity.resolve(guild_id, "klim")
if resolution.ambiguous:
    ...  # ask which member. do not guess

edges = await memory.graph.between(guild_id, x_id, y_id)
shared = await memory.graph.shared(guild_id, x_id, y_id)
left, right = await memory.graph.shared_attributions(guild_id, x_id, y_id)
per_user = await memory.graph.shared_n(guild_id, (x_id, y_id, z_id))
stances = await memory.graph.entity_stances(guild_id, "movies")
audience = await memory.graph.users_for_entity(guild_id, "golf")
neighbors = await memory.graph.neighbors(guild_id, x_id, depth=2)
similar = await memory.graph.similar_users(guild_id, x_id)

Names go through mention ID, username, display name, then saved real name. If more than one member matches, you get ambiguous instead of a guess.

from icelake import CommandAction

command = await memory.classify_command("hey bot remember that I hate pineapple")
# UserMemoryCommand(action=CommandAction.REMEMBER, target_text="that I hate pineapple", ...)
if command.action is CommandAction.REMEMBER:
    await memory.facts.remember(
        guild_id=guild_id,
        subject_id=user_id,
        text=command.target_text,
        actor_id=user_id,
    )

Third-party facts can carry a relation:

from icelake import ProposedRelation, RelationVerb

await memory.facts.remember(
    guild_id=guild_id,
    subject_id=bob_id,
    text="carol called bob a sore loser during game night",
    actor_id=carol_id,
    speaker_id=carol_id,
    relations=(
        ProposedRelation(verb=RelationVerb.CALLED_OUT, from_token=carol_id, to_token=bob_id),
    ),
)

Opt-out and purge:

await memory.admin.set_opt_out(guild_id, user_id, True)
await memory.admin.purge_user(guild_id, user_id, dry_run=False)

Opt-out applies to both observe and recall.

Typed vocabulary

Every closed set of values is a StrEnum exported from the package root. No magic strings, no guessing. Enum members are plain strings, so they compare equal to and serialize as their values:

from icelake import (
    AliasSource,  # identity.register_alias(source=...)
    AttributionType,  # facts.remember(attribution=...)
    ChannelName,  # RecallQuery.channels / channels(...)
    CommandAction,  # classify_command results
    EntityKind,  # ProposedEntity(kind=...)
    FactCategory,  # facts.remember(category=...)
    FactHistoryKind,  # facts.history() entries
    FactScope,  # FactRecord.scope: USER vs SERVER facts
    HealthStatus,  # ops.health() component states
    IgnoreReason,
    RejectReason,
    ObserveStatus,  # observe receipts
    MemoryTier,  # FactRecord.tier, GuildStats.by_tier keys
    MeterPurpose,  # the library's own LLM call purposes
    MessageRole,  # LlmMessage.role
    NodeType,
    Polarity,
    RelationVerb,  # graph edges
    RecallWarning,  # recall / prompt_context warnings
    RerankerProvider,  # retrieval.reranker
    Scope,  # RecallQuery.scope (retrieval-side only)
    SourceRole,  # citation roles
    StorageBackend,  # config.storage.backend
)

Two vocabularies are intentionally open and accept plain strings alongside the enum: RelationVerb (extraction may produce verbs outside the known set, and unknown verbs are polarity-neutral) and meter purposes (charge your own LLM calls under your own names). Note Scope (retrieval) and FactScope (storage) are different sets. Don't use one where the other is expected.

How extraction works

observe writes the message and enqueues it. A worker claims a batch per guild + author under an exclusive attempt lease. Expired work can be retried; processing is at least once. Short chatter is skipped with no LLM call. Otherwise the pipeline mints roster tokens (p0, p1, server), asks the model for JSON, runs quality gates, and stores what survives.

The model never sees Discord snowflakes. Identity fields may only use tokens minted for that batch. Anything else is dropped. Stored text uses display names. The owner of a fact is the snowflake on subject_id, so a rename adds an alias instead of moving the row.

Invalid JSON is repaired once, then dead-lettered. Contradictions invalidate or supersede the old fact. Nothing is deleted. Profile digests regenerate after extraction.auto_consolidate_after_adds new facts (default 5).

Recall does not call the LLM. Typical queries:

Question Call
what do you know about X identity.resolve("X") then facts.list_for_subject(x)
what does X think about Y graph.between(x, y)
what do X and Y share / disagree on graph.shared_attributions(x, y)
what do X, Y, Z all share graph.shared_n((x, y, z))
who likes movies graph.users_for_entity("movies") (stances + mentions) or graph.entity_stances("movies") (edges)
people connected to X graph.neighbors(x, depth=2)
everything about an entity in context prompt_context(..., entity_hint="golf")

What gets stored

Everything hangs off facts. A fact is one sentence about a user or the server, stored in dm_facts against Discord IDs. The other tables exist to find facts and connect them.

Table What it is Why it exists
dm_facts The memory itself, one row per fact Read by everything
dm_vectors One embedding per fact Semantic recall: "what has he been up to creatively" finds the SoundCloud fact
dm_entities Named things that are not Discord users (places, concepts, orgs, off-server people) 30 facts about "the HOA" cluster on one node instead of scattered text
dm_entity_aliases Surface name to entity slug "Koji Sushi" and "koji-sushi" resolve to the same node
dm_links One row per fact per node it touches "Facts about the asker" is an index lookup, not a text scan
dm_relations Typed edges between nodes, like x -friend_of-> y Powers graph.between, entity_stances, neighbors

Facts have a subject_id (the user the fact is about) or a null subject with scope="server" for guild-wide facts. Vectors, links, and relations all reference fact IDs, so purging a user cascades through every table.

Links vs relations is the part people mix up. A link only says "this fact touches this user or entity". It carries no meaning by itself. A relation says "these two nodes have a typed relationship" and carries a verb, polarity, weight, and the fact IDs that support it. Links come from every stored fact. Relations only exist when extraction or a manual facts.remember call found an actual relationship.

Entities vs aliases: the entity is the node, the alias is how you find it. dm_entities holds one row per thing (slug, display name, kind, and linked_user_id when the entity turns out to be a guild member). dm_entity_aliases maps names to that slug, so recall can go from a word in the message to every fact about the thing.

Recall runs several channels over these tables (vector, keyword, links, baseline, entity, graph hop) and fuses the results. Facts are the payload. The rest are indexes over them.

API

memory.observe(event)
memory.observe_many(events)
memory.flush(guild_id=...)
memory.extract_now(event)  # observe + flush in one step
memory.register_bot_id(bot_user_id)  # never stored as a subject

memory.prompt_context(...)
memory.recall(RecallQuery(...))
memory.cite(reply_text, ctx)  # weave jump links into a finished reply

memory.facts.remember / get / get_all / update / forget / reinforce / history / list_for_subject / search
memory.identity.resolve / register_alias / handle_member_rename / aliases_of / display_name
memory.graph.between / relations_of / entity_stances / users_for_entity / neighbors / similar_users / shared / shared_attributions / shared_n
memory.admin.set_opt_out / get_opt_out / purge_user / export_guild / import_guild / backup
memory.ops.run_pending / recover / backfill_aliases / retry_dead_letters / meter_snapshot / health
memory.events.subscribe(BatchCompleted, handler)
memory.classify_command(text)
memory.regenerate_summaries(guild_id)
memory.stats(guild_id)

discord.py

# pip install icelake[discord]
from discord.ext import commands
from icelake import MemoryConfig
from icelake.integrations import setup_discord_memory

config = MemoryConfig(
    storage="sqlite:///bot-memory.db",
    llm="openai://$KEY@openrouter.ai/api/v1?model=z-ai/glm-5.3-flash&reasoning=low",
)


class MyBot(commands.Bot):
    async def setup_hook(self) -> None:
        self.memory = await setup_discord_memory(self, config)

This wires on_message to observe, refreshes aliases on on_member_update, registers the bot's user id on ready, and adds a /memory command group (me, remember, forget). For a bigger command surface see examples/omni_style_bot.py.

Configuration

MemoryConfig(
    storage="sqlite:///memory.db",  # or mongodb://... with [mongo]
    llm="openai://$KEY@openrouter.ai/api/v1?model=...",
    embeddings="hashing",  # default. see below
    # embeddings="local",
    # embeddings="openai://$KEY@openrouter.ai/api/v1?model=openai/text-embedding-3-small",
    batching={"batch_size_messages": 10, "max_age_seconds": 300},
    retrieval={
        "default_token_budget": 2400,
        # Optional L2 after hybrid RRF. Default is "none" (zero-LLM hot path).
        # "openai://$OPENROUTER_API_KEY@openrouter.ai/api/v1?model=voyageai/rerank-2.5-lite"
        # "local" needs icelake[local-embeddings]. Failure degrades to hybrid order.
        "reranker": "none",
        "hop_depth": 2,  # user -> entity -> other attributions
    },
    budgets={"guild_daily_prompt_tokens": 200_000},
    privacy={"store_raw_messages": True},
    workers={"enabled": True, "count": 2},
)

Useful LLM URL query params: model, temperature=none (omit sampling on reasoning endpoints), reasoning=low, max_tokens, max_tokens_key=max_completion_tokens (Azure), structured_outputs=json_object if the endpoint cannot enforce json_schema. A capability mismatch raises LlmCapabilityError.

postgresql:// is recognized and rejected. There is no Postgres adapter yet. Unknown config keys raise immediately.

Embeddings

Provider Spec Notes
Hashing (default) "hashing" Free, deterministic, no extra deps. Fine for tests.
Hosted "openai://$KEY@openrouter.ai/api/v1?model=openai/text-embedding-3-small" Use this for a real bot.
Local "local" pip install icelake[local-embeddings]

Embeddings select semantic recall and reconciliation candidates. They cannot prove that two claims mean the same thing; reconciliation makes that judgment. The default hashing embedder is n-gram feature hashing, not a neural model. "Loves coding in Go" and "Enjoys programming in Go" usually do not collide, so you get duplicate memories. Switch to hosted or local embeddings for anything you actually run.

Changing embedder invalidates existing vectors. Re-embed or start a new database.

Workers

Setup Config
One process defaults, workers run as background tasks
Bot and worker split bot: workers={"enabled": False}, worker: same DB, loop on await memory.ops.run_pending()
Cron workers off, call ops.run_pending from your scheduler
Several processes share one database, keyed leases keep workers from overlapping

Custom backends

from icelake import ChatLLM, DiscordMemory, Embedder, MemoryStore

memory = DiscordMemory(
    config,
    store=MyPostgresStore(),  # MemoryStore (+ optional .queue / .vectors)
    llm=MyLLM(),  # ChatLLM
    embedder=MyEmbedder(),  # Embedder
    clock=FakeClock(...),
)

The override kwargs are typed (MemoryOverrides): your editor will autocomplete store, queue, vectors, embedder, reranker, meter, llm, small_llm, clock, and id_gen, all checked against the port protocols exported from the package root. Passing llm=None, embedder=None, or reranker=None explicitly disables that capability (degraded mode).

A new store has to pass tests/integration/test_store_conformance.py.

Development

uv sync --group dev
uv run pytest tests/ -q --cov=icelake
uv run ruff check src tests && uv run ruff format --check src tests
uv run mypy

Coverage floor is 90%. mypy is strict.

User-facing changes need a changelog fragment in changelog.d/. See CHANGELOG.md. Releases are cut from the Release workflow. See docs/RELEASE.md.

Status

  • Storage: SQLite (default), MongoDB ([mongo]), in-memory (tests). Postgres is not implemented and postgresql:// fails with a clear error.
  • Default hashing embeddings are not good enough for production recall. See Embeddings.
  • Bad extraction JSON is repaired once, then dead-lettered. Retry with ops.retry_dead_letters.
  • Caps and TTL drop weakest facts first (manual/CORE last). Budget counters are store-backed, so they are correct across N worker processes. Observability meters stay per-process.
  • similar_users is capped Jaccard over entity adjacency.

License

MIT

Production reliability and migration

Authoritative fact, graph, input-receipt and projection-job writes commit together. SQLite uses native transactions and synchronous=FULL. MongoDB now requires a replica set or a transaction-capable cluster and PyMongo 4.17+; standalone MongoDB can no longer silently claim successful multi-document commits. Custom stores must implement real transaction() semantics and the projection-job methods in MemoryStore. In-memory storage remains an ephemeral reference adapter. Mongo uses the driver's bound sessions and transaction support.

Input delivery is at least once. A durable receipt for each message and processing scope makes committed batches replay-safe, including re-batched input after an acknowledgment failure. Receipts contain identifiers/hashes, not conversation text; they persist indefinitely to prevent old backfills from being counted again. Plan revisions are checked before mutations, and stale plans retry from current memory. Event hooks are best-effort notifications after commit, not a durable external event bus.

Vectors are a rebuildable projection. A transactional outbox survives crashes and independent index outages; old revisions cannot overwrite newer vectors. Built-in workers drain bounded recovery work, and external schedulers call await memory.ops.recover(). Failures defer individual jobs so one unavailable fact does not block the rest. ops.health() reports pending indexing as degraded. An outage can temporarily reduce semantic recall while committed facts remain available through other retrieval channels. Injected vector adapters must honor VectorItem.version with an atomic conditional upsert.

config = MemoryConfig(
    storage="sqlite:///memory.db",
    recovery={
        "max_attempts": 3,
        "retry_initial_seconds": 5,
        "retry_max_seconds": 300,
        "projection_batch_size": 32,
        "projection_timeout_seconds": 30,
    },
    access={"require_audience": True, "timeout_seconds": 3},
)

Batch retry state and jittered backoff survive restarts. Community-window retries also stop after the configured attempt limit; the same operator retry API resets their budget. Community roll-ups still use the configured recent-message window and raw-message retention, so they are not an unlimited historical backfill. Exhausted batches remain in dead letters; repair the cause, then use ops.retry_dead_letters(guild_id). Provider retries remain separately bounded by llm.max_retries; the worst-case provider attempt count is the product of both limits. Cancellation retains the input, and crashed leases are reclaimed by ops.recover().

For explicit commands, supply facts.remember(..., idempotency_key=interaction_id). The same key returns the same fact without reinforcing it; conflicting reuse is rejected. A retry cannot recreate a fact erased after the original command. Without a key, repeated remember/reinforce calls intentionally count as new input.

Retention and learning

Durable personal facts, preferences, interests, relationships and experiences have no automatic expiry by default. General observations last 365 days; explicit near-term plans retain short deadlines. Recency changes retrieval rank without deleting a quiet member's profile. Corrections, source deletion, consent and per-subject capacity limits still apply. Age-based retirement is opt-in via lifecycle.forget_retention_floor; long_term_days=None means no expiry.

Existing stored deadlines are not silently rewritten. After a backup, icelake.lifecycle.tiers.retention_extension can plan extensions for active, unexpired automatic facts; apply with a version check. Deleted, expired and superseded memories are never revived by that helper.

This follows the separation of evolving facts and retrieval in Graphiti and optional expiry in Mem0. It is not an implementation-equivalence claim about proprietary assistants.

Channel permissions

recall and prompt_context now require an Audience by default. Missing context returns no facts and an AUDIENCE_REQUIRED warning. Build it from trusted transport metadata, never model output or unchecked request parameters. The reply audience means everyone who can read the destination, not only the person asking.

The transport-neutral default restricts disclosure to the same source channel; the embedding application must authenticate the caller and authorize the audience. For Discord, setup_discord_memory installs DiscordChannelAuthorizer, which reads fresh roles, channel overwrites and membership through REST, checks private-thread membership, and abstains if history visibility cannot be proven for all readers. Role grants work even when @everyone has no history permission. Straightforward role policies need no member scan; otherwise the adapter checks a complete REST snapshot (at most 10,000 members, requiring the Members intent). Cross-channel reads require source access for every destination reader. Private thread audiences are conservatively bounded by their parent channel's readers. Requests are timeout-bounded and API failures deny access. Retrieval reports ACCESS_RESTRICTED or ACCESS_UNAVAILABLE; strict visible_facts and visible_sources calls raise MemoryAccessError instead of implying no memory exists. This uses discord.py's channel permission calculation.

context = await memory.prompt_context(
    guild_id=guild_id,
    asker_id=user_id,
    audience=Audience(channel_id=reply_channel_id, requester_id=user_id),
    text=question,
)

All evidence channels must be allowed. Unknown provenance and mixed private/public facts are excluded from protected retrieval. Filtering happens before external reranking and is checked again before returning results. cite() rechecks the context's fact revisions, consent and audience before emitting source links. Multi-hop relation traversal is disabled for protected retrieval because its aggregated edges are not audience-partitioned. Direct evidence-backed retrieval remains available. Guild-wide summaries and mined name aliases are omitted from protected prompts; channel-partitioned extraction and reconciliation prevent private context from being incorporated into a public-channel fact.

facts, graph, identity and admin are trusted control-plane APIs, not end-user authorization boundaries. Do not expose their raw results in public bot replies. /memory me only includes channel-authorized facts plus self-authored manual commands in its ephemeral response. Unscoped manual memories remain available for that private inspection; to expose a manually created fact in a channel, attach genuine source_refs from authorized source messages. access.require_audience=False is an explicit trusted/offline compatibility mode, not a production authorization policy. Reauthorize immediately before sending if your application holds a generated reply for an extended period.

The runnable bot examples install live authorization and pass audience metadata for both generated replies and profile commands. They omit raw channel-history injection and unscoped graph displays. Name lookups use current Discord membership; private memory aliases are not disclosed. Graph inspection remains demonstrated separately in the trusted relationship_queries.py example.

Operational backups and restore drills

admin.export_guild() is a portable content export, not a disaster-recovery backup. SQLite operational snapshots include facts, aliases, incidence links, relations, history, summaries, governance generations, opt-outs, queue contents, leases, commit receipts, budgets, cursors, cached data, vectors and pending work. The adapter uses SQLite's online backup API.

from pathlib import Path
from icelake.adapters.sqlite.backup import restore_backup

await memory.admin.backup(Path("backups/snapshot-2026-10-17"))  # new directory
# Restore BEFORE opening this new database with a client:
await restore_backup(
    Path("backups/snapshot-2026-10-17"),
    Path("recovered.sqlite"),
    max_bytes=10 * 1024**3,
)

Restore verifies schema, size, SHA-256 and SQLite integrity, resets interrupted claims, and atomically publishes a new destination. Existing databases are never overwritten. Checksums detect corruption; they do not authenticate an untrusted archive. Independent queue/vector overrides require coordinated native snapshots or an explicit rebuild plan; admin.backup() rejects that configuration rather than silently producing an incomplete deployment backup.

Mongo deployments can use managed service backups or MongoReplicaBackup from icelake.adapters.mongo.backup. It wraps official mongodump --oplog and mongorestore --oplogReplay --stopOnError, with deadlines, archive size limits, checksums and a private credentials file. Its scope is the entire replica set, including all databases, because a consistent oplog dump cannot be limited to one guild/database. Supply deployment URIs without a database path. Restore only to an empty, offline deployment; a failed restore must be discarded and retried into a fresh target. Native tools must be installed separately. Follow MongoDB's backup requirements and restore version requirements.

Keep application configuration, embedding-model identity, encryption keys and secrets in your existing deployment backup system. Restore with the matching embedding configuration. Reconcile erasures/opt-outs newer than the snapshot before opening the restored system to users, and enforce retention/access control on backups themselves. A snapshot cannot reconstruct changes made after it.

Conversation evaluations

evals/conversations.py replays annotated transcripts through the configured extractor, reconciliation, storage and retrieval. It accepts JSONL ConversationCase records with provenance (consented_redacted or synthetic), a permission reference, timestamped messages, and checkpointed probes. Gold labels include subject IDs, expected text fragments and supporting message IDs; probes also specify forbidden content, an audience and required abstention.

from pathlib import Path
from evals.conversations import evaluate_case, load_cases, save_results
from icelake import StorageConfig, WorkersConfig

cases = load_cases(Path("redacted-conversations.jsonl"))
results = []
for case in cases:
    result = await evaluate_case(
        case,
        lambda clock: DiscordMemory(
            config.model_copy(update={"storage": StorageConfig(url="sqlite://:memory:"),
                                      "workers": WorkersConfig(enabled=False)}),
            clock=clock,
        ),
    )
    results.append(result)
save_results(Path("evaluation-results.json"), tuple(results))

Use a fresh store per case and the supplied replay clock. The harness rejects empty datasets, invalid/future evidence labels, duplicate message IDs, live background workers and incomplete ingestion. Reports include matched gold facts, relevant/returned counts, source attribution, forbidden disclosures, abstention, retrieval latency and provider usage, without raw transcript text. Fragment matching and provenance checks are deterministic diagnostics, not semantic truth judgments or a substitute for human annotation review.

The existing scripted golden suite is a regression suite; its model outputs do not establish extraction accuracy. No real-conversation corpus or proprietary assistant parity is claimed by the bundled synthetic harness tests. Supply consented, redacted conversations and run the chosen production model to measure actual quality and cost.

Metadata

Release files for icelake 0.7.4

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for icelake 0.7.4
File Size Uploaded Provenance Actions
icelake-0.7.4.tar.gz 543.6 kB Without provenance Details for icelake-0.7.4.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for icelake 0.7.4
File Interpreter ABI Platform Actions
icelake-0.7.4-py3-none-any.whl Python 3 none any Details for icelake-0.7.4-py3-none-any.whl

Total release size: 801.9 kB

Release files / icelake-0.7.4.tar.gz

Download URL icelake-0.7.4.tar.gz
Size 543.6 kB
Tags Source
SHA-256 checksum
How to use checksums
44a4f86d1193b6bae13f641dc2d418e177bf5179c22e5f969cbf6e4ee50c87fe
BLAKE2b-256 checksum
How to use checksums
0efc93918da03e211e3dee85886e4b5d30a40dcc9979a612e27e306cbf954059
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via uv/0.13.7 {"installer":{"name":"uv","version":"0.13.7","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"26.04","id":"resolute","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}

Release files / icelake-0.7.4-py3-none-any.whl

Download URL icelake-0.7.4-py3-none-any.whl
Size 258.3 kB
Tags Python 3
SHA-256 checksum
How to use checksums
7d9ec2ebe5020d3c91b1dcdd01417d945fa52419e3cfaf0a24316fc06345588f
BLAKE2b-256 checksum
How to use checksums
4efb0836612111d362f4eda93d4a8ca63d5dc1a602f3e2f8edf9b1b01f357215
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via uv/0.13.7 {"installer":{"name":"uv","version":"0.13.7","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"26.04","id":"resolute","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}

Security for icelake 0.7.4

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page