Athenaeum

Production-tested agentic memory for teams deploying multiple AI agents. Append-only intake, a tiered librarian that compiles raw observations into a trustworthy wiki, and a sidecar that makes recall happen passively on every turn.

Is this for me? If you're running more than one agent on shared knowledge — or if you want agents and humans reading and writing the same institutional memory — yes. If you're building a single-user chatbot, mem0 or Letta may be a better fit.

What is this?

A context window forgets everything the moment a session ends, and even within a session it can't tell a durable fact ("Acme is a client") from a passing remark. Athenaeum is a memory layer that sits outside any one agent session: agents write observations to an append-only intake log, a separate compiler (the librarian) turns that raw stream into a structured, deduplicated wiki of entities, and a sidecar injects the relevant slice of that wiki back into context automatically on every turn. The result is memory that survives across sessions, across agents, and across a team — not just across turns of one conversation.

Full rationale, comparison with alternatives (Claude memory, Anthropic's memory tool, RAG, Karpathy's gist, mem0/Letta/Zep/Cognee), and the lessons from running it on our own operations live in docs/why-athenaeum.md. For the companion blog post: What We Learned Running Our Own Operations on Agentic Memory.

Key features and why they're built this way

The librarian compiles raw intake into a wiki — it doesn't store verbatim

What: Agents can only append to raw/. A separate process, the librarian, is the only writer to wiki/, and it runs the observation through a tiered pipeline — programmatic normalization, a fast classifier that routes the observation to a known or new entity, and a capable model that merges it into the entity page and resolves simple contradictions.

Why: If every agent could edit the wiki directly, three facts about the same person surfaced across three sessions would become three drifting person pages, and a bad write from one agent would be indistinguishable from a good one. Routing everything through one compiler makes safety a property of structure, not of trusting every agent to be a careful writer — the librarian snapshots the wiki to git before every run, so a bad merge is a git revert away.

Contradiction detection escalates to a human decision queue

What: When the pipeline can't confidently resolve a conflict between an incoming observation and the existing wiki, it doesn't guess — it appends a block to wiki/_pending_questions.md for a human to answer. The list_pending_decisions MCP tool (and athenaeum decisions on the CLI) surfaces this queue, unified with pending merge proposals, as one "decisions needed" list.

Why: An automated memory system that silently picks a side on every ambiguous conflict will eventually silently pick the wrong side, and nobody will know to check. Some conflicts are cheap to auto-resolve; the ones that aren't need a human in the loop, and that only works if there's a durable, listable surface for them rather than a conflict resolved by whichever write happened last.

Source precedence is a 9-tier taxonomy, not last-write-wins

What: Every claim carries a source, and when two claims conflict, the resolver picks a winner using a fixed 9-tier precedence order (from docs/why-athenaeum.md and resolutions.py): direct user statement, curated public profile, third-party authoritative API, consensus public source, agent-observed-from-artifact, LLM-generated, pipeline script, model prior, and unsourced last.

Why: "Whichever fact was written most recently wins" is how memory systems quietly go wrong — a stale scraped fact can overwrite something the user said directly. A precedence order means a low-authority source can never silently clobber a high-authority one; genuine ties still fall through to a human question rather than a coin flip.

Recall is scoped by audience, fail-closed

What: athenaeum serve --audience <role> pins a restricted read scope for the life of that server process. Every page-content-bearing tool — recall and every list tool — applies the same predicate: a restricted caller sees only pages it's explicitly authorized for. Untagged pages fail closed for a restricted audience (the owner, with no audience pinned, still sees everything).

Why: A secondary or scheduled agent (an email-drafting routine, say) often needs some of the wiki but must never reach PII or client-confidential pages. Defaulting to "visible unless labeled private" means a missed label leaks; defaulting to "hidden unless labeled visible for this audience" means a missed label is merely annoying, not a leak.

Spend ceilings and a run lock bound the background compiler

What: athenaeum run enforces a per-run API-call budget (ATHENAEUM_MAX_API_CALLS, default 800) and takes a shared run lock (RunLock, issue athenaeum#309) so run, ingest, reindex, and session-end are single-flight against each other on the same knowledge root.

Why: The librarian is an LLM-driven background process with no human watching each call — without a hard ceiling, a bad day (a big backlog, a runaway retry loop) turns into an unbounded bill. And because the librarian mutates the wiki with git commits, two overlapping runs writing at once is a correctness bug, not just a cost one — the lock guarantees a single writer.

A provider abstraction lets the pipeline run on the API or a Claude Code subscription

What: All LLM calls go through one factory (athenaeum.provider) with two backends: api (the Anthropic SDK, metered) and claude-cli (drives the operator's own claude binary against their Code subscription, no API key). The api backend wraps anthropic.Anthropic with parameters passed through unchanged, so behavior is byte-for-byte identical to the pre-abstraction code; the claude-cli backend mirrors the same call surface so callers need no branching logic.

Why: Not every operator wants to meter a separate API key for a nightly batch job when they already pay for a Claude Code subscription — but the tiers, prompts, and call sites shouldn't need two implementations to support both. One seam, two transports, identical prompts either way.

Storage adapters are an extension point (not yet used in-tree)

What: storage.py resolves each wiki entity class to a storage surface — a backing store plus a corpus policy — through a small registry (register_adapter / storage.adapters in config). Two adapters ship built-in: the default wiki-markdown-embedded surface, and an excluded surface (all-false corpus policy) that PII/archival content is routed through so it's excluded from embed/recall/merge by construction.

Why (honestly scoped): This is presented as an extension point, not a finished feature: only one corpus-policy bit (is_merge_eligible) currently has a real call site in src/ (wiki_dedupe.py); the is_embedded and is_recallable bits exist on StorageAdapter but nothing in src/ reads them yet, and there's no third-party adapter registered in-tree today. The seam exists so a future storage surface (e.g. a database-backed wiki) is a config change, not a core rewrite — but as of this writing it has one real built-in consumer (the PII-exclusion surface), not a general policy engine.

Note on reasoning tiers. reasoning_tiers.py implements a tiered reasoning pipeline that screens merge proposals before the human queue (DEFAULT_TIER_CHAIN, the pipeline's own default chain, is the empty tuple), but it is not part of any default run path. It has two production call sites in merge.py — t1_screen_rejects_merge_proposal (reject/pass-up only) and t2_screen_merge_proposal (which can auto-apply a safe-class merge with no human review) — both gated behind the same resolve_reasoning_tier_auditing_enabled flag, an explicit opt-in that defaults to off. Until an operator turns that setting on, production merge behavior is unaffected. Full writeup, including when it's worth turning on and what it costs: Reasoning-tier screening (T1/T2).

The MCP surface

Athenaeum ships an MCP server (athenaeum serve) exposing 11 tools so AI agents can write to raw intake, search the compiled wiki, and triage the human-decision queue — 7 read-only, 4 that mutate human-decision state.

The three decision-queue mutators (resolve_question, resolve_merge, review_audit_item) plus every page-content-bearing read tool are subject to audience scoping: a restricted (--audience) server process fails those mutators closed entirely, since adjudicating the operator's decision queue is owner-only. See docs/security-posture.md.

pip install 'athenaeum[mcp]'
athenaeum serve --path ~/knowledge

# Smoke test the round-trip without a live session
athenaeum test-mcp

Claude Code integration. Add to your MCP config and it auto-starts with every session:

claude mcp add --scope user athenaeum -- athenaeum serve --path ~/knowledge

Custom raw/wiki locations. The raw and wiki roots default to <path>/raw and <path>/wiki. To point them at independent locations — for example a read-only mounted wiki, or an existing config that predates this command — set KNOWLEDGE_RAW_PATH and/or KNOWLEDGE_WIKI_PATH; each overrides its root individually while --path remains where athenaeum.yaml and extra intake roots resolve:

KNOWLEDGE_RAW_PATH=/data/knowledge/raw \
KNOWLEDGE_WIKI_PATH=/data/knowledge/wiki \
  athenaeum serve --path ~/knowledge

Scoped read access (secondary agents). By default athenaeum serve exposes the whole wiki to recall — the owner sees everything. If you wire a secondary/scheduled agent (e.g. an email-drafting routine) to this server, pin it to a restricted audience so it can reach operational pages but never your PII / client-confidential / financial ones:

# This server process may only recall pages an "ops" audience is allowed to read.
athenaeum serve --path ~/knowledge --audience ops

The audience is pinned by the operator at serve time and cannot be widened by the caller. Pages carry access: (open/internal/confidential/personal) and/or an audience: role list; untagged pages fail closed for a restricted audience (owner still sees them). This is a single-owner read filter, not a multi-user ACL. See docs/security-posture.md and docs/configuration.md for the frontmatter model and the fail-closed enforcement.

Write-time PII screening (opt-in, off by default). Complementing the read-time audience filter above, an intake screener can classify each remember() payload before it is stored and auto-label matched medical content access: personal so a restricted recall never surfaces it (it labels, never drops). It is off by default — a deliberate choice, not a gap: unscreened intake already defaults to the non-world-readable access: internal, so nothing is world-readable unless a page is explicitly labeled open. Turn it on in athenaeum.yaml:

screening:
  medical:
    action: label_restrict   # default: off

Example round-trip:

User: Tristan's partner is Amanda; they met at Stanford GSB.

(Claude calls remember(content="Tristan's partner is Amanda; they met at Stanford GSB.", source="claude-session"))

A raw observation lands in raw/claude-session/20260417T…-…md. On the next athenaeum run, the pipeline compiles it into Tristan's wiki entity (under "Key Contacts") and Amanda's own entity if she doesn't exist yet. Later sessions can ask "who is Amanda?" and recall returns the compiled page.

Installation

Requires Python 3.11+.

pip install athenaeum

Athenaeum ships an optional reasoning-tier screen (T1/T2) that can pre-screen merge proposals before they reach your human review queue — off by default, since the T2 tier can auto-apply a merge with no human review. Worth turning on once your merge queue outgrows manual triage; see Reasoning-tier screening (T1/T2) for when and how.

Quick start

# Initialize a knowledge directory
athenaeum init                  # default: ~/knowledge
athenaeum init --path ~/my-knowledge

# Run the librarian (compile raw intake → wiki entities).
# `athenaeum run` needs ANTHROPIC_API_KEY — use --dry-run to explore keyless.
# `run` and `status` operate on ~/knowledge by default (point elsewhere with
# `--path`; `run` also accepts `--knowledge-root` as the original spelling).
athenaeum run
athenaeum run --dry-run         # inspect without writing

# Check status
athenaeum status

Full run with custom paths and budgets (--max-api-calls 200 here deliberately lowers the per-run API budget below the default of 800 — omit the flag to accept the default):

athenaeum run \
  --raw-root ~/knowledge/raw \
  --wiki-root ~/knowledge/wiki \
  --path ~/knowledge \
  --max-files 50 \
  --max-api-calls 200 \
  --verbose

On-demand ingest & reindex (issue athenaeum#349)

The nightly athenaeum run is the batch path. When an agent (or you) needs a just-remembered fact to become recallable now — decoupled from the nightly cadence — use the two on-demand commands. Both are single-flight (they share the same run lock as run, issue athenaeum#309) and print a one-line JSON summary with counts and duration; both exit non-zero on failure.

# Compile only raw intake that is new/changed since the last ingest, then
# refresh the search index — the round-trip that makes a memory recallable.
athenaeum ingest              # --incremental is the DEFAULT (fast no-op if none)
athenaeum reindex             # --incremental hash-diff delta (depends on athenaeum#348)

athenaeum ingest --full       # recompile all pending raw intake
athenaeum reindex --full      # rebuild the index from scratch
athenaeum ingest --session <id>   # scope new/changed detection to one session

ingest --incremental tracks a content-hash stamp (~/.cache/athenaeum/ingest-manifest.json, mirroring the athenaeum#348 index manifest), so it is a fast no-op when nothing has changed. tier0_passthrough pre-structured intake compiles with no LLM cost. reindex is the canonical name; rebuild-index remains as a back-compat alias for the exact same command. The reusable engine lives at athenaeum.librarian.ingest (the SessionEnd path, issue athenaeum#350, calls it directly).

Cross-agent same-day recall — session-end (issue athenaeum#350)

remember writes only to raw/; recall reads only the compiled wiki/ index; the librarian that compiles raw/→wiki/ runs nightly. Without intervention a memory written by one agent at 10:00 is invisible to every other agent until the next nightly run — a ~24h gap. athenaeum session-end closes it by composing the two on-demand steps above into one change-gated command the SessionEnd hook (and the nightly-after-librarian path) invokes:

athenaeum session-end                     # incremental ingest + reindex (DEFAULT)
athenaeum session-end --session <id>      # scope new/changed detection to one session
athenaeum session-end --full              # force a full recompile + full index rebuild
athenaeum session-end --dry-run           # cheap manifest-diff preview — no compile, no reindex, no model load

Both steps are change-gated so an idle SessionEnd is cheap:

1. Incremental ingest of this session's new/changed raw intake — a fast no-op (zero LLM) when nothing is new; tier0 structured entries compile with no model cost.
2. Then reindex, but only when the compile actually ran and succeeded. An idle SessionEnd (no new raw) does no LLM work and no reindex; a failed compile never indexes a half-built wiki; a --dry-run touches nothing.

The result is a memory remembered in session A becoming recallable — as a fully-resolved wiki entry — in session B the moment A ends, no waiting for the nightly librarian. It is single-flight (shares the run lock, athenaeum#309) and prints a one-line JSON summary nesting the ingest counts plus the reindex page count. The reusable engine is athenaeum.librarian.session_end; the CLI is a thin wrapper. The hook that fires this at SessionEnd lives in your Claude Code workspace config — this repo ships the command it calls.

Maintenance & inspection commands

Two subcommands operate over the compiled wiki outside the nightly run loop — one read-only, one with an opt-in destructive --apply:

# Read-only: find claims restated across distinct wiki entities.
athenaeum claims --find
athenaeum claims --find --threshold 0.9 --path ~/knowledge

claims --find is a cross-entity recurring-claim detector. It scans the wiki, embeds each claim via the recall-index embedding provider, and prints a YAML report grouping claims that recur across two or more distinct entities (default cosine cutoff 0.85, override with --threshold). It never mutates wiki/ — it only reports. With no embedding backend available it degrades to an empty report rather than failing.

# Cluster compiled wiki pages against each other and propose merges.
athenaeum dedupe wiki-pages
athenaeum dedupe wiki-pages --dry-run --threshold 0.6

dedupe wiki-pages clusters already-compiled concept/reference/principle wiki/*.md entity pages by topic/embedding similarity (issue athenaeum#290) — complementing the raw-intake clustering that runs during athenaeum run. True duplicates are routed through the existing wiki/_pending_merges.md / resolve_merge approval flow (never auto-applied). Writing a proposal is idempotent — rerunning for a source set already proposed is a no-op. --dry-run previews without writing; --threshold overrides librarian.cluster_threshold (default 0.55, the same cutoff the raw auto-memory cluster pass uses). athenaeum run also runs this pass automatically whenever wiki/ exists — there is currently no toggle to opt out of the run-embedded pass; failures are logged (wiki-page dedup pass failed; continuing run) and non-fatal to the run.

# Archive resolved (approved/rejected) blocks out of the live sidecar.
athenaeum ingest-merges --path ~/knowledge

resolve_merge does not archive on its own — like resolve_question, it only flips the checkbox in place. Run ingest-merges (issue athenaeum#299) to move every resolved block out of wiki/_pending_merges.md into wiki/_pending_merges_archive.md (newest-first, append-only), keeping the live sidecar limited to genuinely open proposals. Idempotent — this must be scheduled (or run periodically); nothing else archives resolved merges, which is exactly how the live file grew to 5MB/67K lines in production before this command existed (athenaeum#299).

# Dry-run (default): print the kill-list + retained-list, change nothing.
athenaeum auto-memory prune
# Apply: git rm the kill-list in one commit and rebuild the recall index.
athenaeum auto-memory prune --apply

auto-memory prune retires operational/ephemeral wiki/auto-*.md pages (throwaway scratch scopes, install-token boilerplate, and pages flagged ephemeral: true) using the same classifier the intake gate applies. Dry-run is the default: it prints the kill-list and retained-list with a reason per page and exits 2 when candidates exist (a CI / sign-off signal), 0 when there is nothing to prune. --apply git rms the kill-list in a single labeled commit and then rebuilds the recall index; the commit pathspec is scoped to the kill-list, so unrelated staged work is never swept in. Like move-then-retire, recovery is git-only — the command refuses to run outside a git repo and never hard-unlinks, so a pruned page is recoverable from history.

Answering pending questions

When Tier 3 can't resolve an ambiguity or a principled contradiction, the librarian escalates to wiki/_pending_questions.md. Each escalation lands as a block like:

## [2026-04-20] Entity: "Acme Corp" (from sessions/20240406T120000Z-aabb0011.md)
- [ ] Is Acme still Series A after the 2026 recapitalisation?
**Conflict type**: principled
**Description**: Prior wiki says Series A; the 2026-04 raw file implies Series B.

You resolve a question one of two ways — pick whichever fits your workflow:

Option 1 — Edit the file directly

Flip [ ] to [x] on the checkbox line and type your answer below the checkbox (above or below the conflict-type / description lines — either works; the parser strips those metadata lines when extracting the answer):

## [2026-04-20] Entity: "Acme Corp" (from sessions/20240406T120000Z-aabb0011.md)
- [x] Is Acme still Series A after the 2026 recapitalisation?

They closed Series B on 2026-03-12, led by Acme Growth Partners.
The 2026-04 raw file is correct; the prior wiki entry is stale.

**Conflict type**: principled
**Description**: Prior wiki says Series A; the 2026-04 raw file implies Series B.

Option 2 — Use the MCP tools

For containerized agents that can't touch the filesystem, athenaeum serve exposes the full 11-tool surface documented above, including:

• list_pending_questions() returns unanswered blocks as JSON — each item carries a stable id derived from the header + question text.
• resolve_question(id, answer) flips the checkbox and writes the answer body under it. It does not archive on its own — archival runs on the next ingest-answers pass.
• list_pending_decisions() returns the unified queue — pending questions AND resolver merge proposals in one call, each tagged type: "question" | "merge" (issue athenaeum#401). Merges name their source pages by human title with a one-line gist so the item reads as an answerable question.

One unified "decisions needed" list

Athenaeum accumulates two human-decision queues: questions (contradiction detector) and merges (resolver merge proposals). athenaeum decisions unifies both so there is one place to look:

athenaeum decisions count            # "7 decisions pending (3 questions, 4 merges; oldest 30d)"
athenaeum decisions list --json      # both queues, each tagged type, oldest first
athenaeum decisions next             # the single oldest decision

Each merge item is rendered as an answerable question — the source pages are named by their frontmatter name: (not the uuid-slug) with a one-line gist each — because cosine topic-similarity alone is not "should-merge" and misleads without the pages' own words. The merges half is also available on its own via athenaeum merges {list,next,count,revalidate,provenance} (the mirror of athenaeum questions, plus two merge-only modes):

athenaeum merges list  [--limit N] [--json]        # all unresolved proposals
athenaeum merges next  [--json]                     # the oldest unresolved proposal
athenaeum merges count [--json]                     # "N unresolved (oldest: <iso-date>)"
athenaeum merges revalidate [--apply] [--json]      # re-check the queue against the CURRENT gate
athenaeum merges provenance [--canonical-slug S] [--merge-id ID] [--json]

revalidate is the first move for "is the merge queue healthy?" It re-runs the current suppression gate (size cap + confidence floor) against every unresolved proposal and reports each one's n_sources and the suppression reason for anything that would now be rejected. Proposals queued before the gate tightened (issue athenaeum#400/#421) don't get re-checked on their own — this command is how you find them. It is dry-run by default; pass --apply to archive the stale ones to wiki/_pending_merges_archive.md (non-destructive — moved, never deleted). provenance is the read side for merges that already executed: which source pages a completed merge relied on, from wiki/_merge_provenance.jsonl (issue athenaeum#425).

Never hand-parse wiki/_pending_merges.md. It is a hand-rolled markdown sidecar with nested code fences and multi-line fields — grep/awk against it is fragile and has burned real investigation time. parse_pending_merges() (src/athenaeum/pending_merges.py) is the only sanctioned reader; the athenaeum merges subcommands above are the sanctioned CLI surface built on top of it. Also note: the MCP list_pending_decisions / resolve_merge view adds derived fields that do not exist in the file, most visibly sources_omitted (a rendered-vs-total source count computed in decisions.py) — don't assume a field you saw in that view is present in the markdown or in athenaeum merges JSON output.

Step 2 — ingest the answers

Either way, run:

athenaeum ingest-answers --path ~/knowledge

Each [x] block is rewritten as a raw intake file under raw/answers/{timestamp}-{entity-slug}.md with frontmatter linking back to the original source, then moved into wiki/_pending_questions_archive.md (newest-first, append-only — answered blocks are never deleted, only moved). The next athenaeum run picks the raw file up like any other intake and folds the answer into the wiki entity.

Re-running with no new [x] blocks is a no-op. Malformed blocks are preserved in place and logged to stderr, so a corrupt single entry cannot poison the rest of the file.

Transparent sidecar (Claude Code hooks)

For a fully passive experience where Claude auto-recalls relevant context on every prompt and saves observations without explicit commands, configure Claude Code hooks:

1. Copy the example hooks from examples/claude-code/ to your scripts directory.
2. Add hook entries to ~/.claude/settings.json (see examples/claude-code/settings-snippet.json).
3. Add CLAUDE.md instructions for proactive memory (see examples/claude-code/CLAUDE.md.example).

This gives you:

• Auto-recall — an FTS5 index is built at session start (~300ms); each user message triggers a <50ms search that injects relevant wiki pages into context.
• Auto-remember — Claude proactively saves important facts without being asked.
• Context checkpointing — observations are saved before context-window compaction.

Full setup guide, smoke test, and environment-variable reference: examples/claude-code/README.md.

Building your own adapter

Any external source — an API, an export file, a message feed, a scraper — can feed Athenaeum by writing raw-intake files that the librarian compiles into the wiki. That seam is a small, stable contract: a source only appends a raw file; the librarian is the only writer to the wiki.

• The contract — docs/adapter-contract.md: file location, frontmatter shape, provenance, idempotency, and how compilation reconciles duplicates/updates.
• Guided walkthrough — the bundled adapter-authoring skill (ships in the package under skills/) teaches an agent or a human how to build a custom adapter step by step.
• Worked example — examples/adapters/minimal_adapter.py: a synthetic, runnable adapter you can copy as a starting point.

Integrations

• Claude Code auto-memory — bridge ~/.claude/projects/<scope>/memory/ into Athenaeum's raw/ intake so the librarian can cluster, merge, and contradiction-check Claude Code's durable memory alongside other sources. A complete worked adapter for the auto-memory intake lane. See docs/integrations/claude-code.md.
• Contradiction detection — pipeline overview, cross-scope modes, source-precedence taxonomy, configuration reference, and cost model for the auto-memory contradiction path. See docs/contradiction-detection.md.

Vector search (optional)

Athenaeum supports a vector search backend (chromadb + all-MiniLM-L6-v2) for semantic recall alongside the default FTS5 keyword backend. The recall hook runs a hybrid FTS5 + vector merge when vector is configured — each backend rescues a failure class the other has (short-query proper-noun collisions for vector; no-lexical-overlap semantic queries for FTS5).

pip install 'athenaeum[vector]'

Enable it in athenaeum.yaml:

search_backend: vector

Full walkthrough and the four invariants a future simplification must not remove: docs/recall-architecture.md.

Query-topic extraction (optional)

athenaeum query-topics "your prompt" runs a Haiku classifier that returns substantive topics and ignores meta-instructions:

$ athenaeum query-topics "Without calling any tools, quote the block about Return Path verbatim"
Return Path

The naive regex+stopword fallback returns block,calling,quote,return,tools,verbatim,without — burying "Return Path" behind meta-instruction tokens. The example recall hook uses query-topics to rescue named-entity recall on instruction-heavy prompts and falls back silently to the regex extractor if the API key or CLI is unavailable.

Environment variables

The table below covers the common knobs. The exhaustive list — every env var, yaml key, and CLI flag with its code default and precedence chain — lives in docs/configuration.md.

Shell-env overrides. AUTO_RECALL and SEARCH_BACKEND are read from the shell environment after the hook sources ~/.cache/athenaeum/config.env, so exports in your shell profile beat the cached config. Intentional (lets you A/B-test a backend without editing athenaeum.yaml), but it's the first thing to check when the hook "ignores" a config change.

Claude Code auth caveat. Claude Code's own CLAUDE_CODE_OAUTH_TOKEN is scoped to its inference endpoint, and the Anthropic Messages API rejects it with 401 OAuth authentication is currently not supported. The pipeline and example hooks need a separate console API key — see docs/recall-architecture.md for the 1Password bootstrap pattern.

Configuration

Settings are resolved in the order CLI flag > env var > <knowledge_root>/athenaeum.yaml > built-in default, so a one-off shell export beats the yaml without requiring an edit. The canonical reference for every knob — librarian budgets, model selection, contradiction/resolver tuning, recall/search, and hook environment — is docs/configuration.md. As one example, the resolver model lives under the top-level models: block, and the rest of the resolver's behavior knobs live under resolve::

models:
  resolve: claude-opus-4-7        # ATHENAEUM_RESOLVE_MODEL

resolve:
  auto_apply: true                # ATHENAEUM_RESOLVE_AUTO_APPLY (default: true)
  auto_apply_threshold: 0.90      # ATHENAEUM_RESOLVE_AUTO_APPLY_THRESHOLD, [0.0, 1.0]
  full_body_token_cap: 1500       # ATHENAEUM_RESOLVE_FULL_BODY_TOKEN_CAP, per-side body cap (~4 chars/token)

When auto_apply is on and a proposal's confidence meets or exceeds auto_apply_threshold, the pending-question block is auto-flipped to answered with an Auto-resolved: true audit-trail tag. See docs/auto-resolve.md for the full lane, including how to disable, lower the threshold, or reverse an auto-resolution.

Alternative model gateways. All model calls go through the Anthropic SDK, which honors ANTHROPIC_BASE_URL — so a LiteLLM proxy or any Anthropic-compatible gateway can serve alternative models with zero code change. Only Claude models are first-party tested; see docs/configuration.md for the details and #234 for multi-provider tracking.

Data formats

Raw intake lives in raw/{source}/*.md with the naming convention {timestamp}-{uuid8}.md (e.g., 20240406T120000Z-aabb0011.md). Each file is a plain markdown document containing observations, notes, or session transcripts. The {source} directory identifies the origin (e.g., sessions, imports).

Wiki entity pages live in wiki/ with YAML frontmatter:

---
uid: a1b2c3d4
type: person
name: Alice Zhang
aliases: [Alice]
access: internal
tags: [active]
created: '2024-04-06'
updated: '2024-04-06'
---

Entities are indexed in wiki/_index.md grouped by type. Conflicts requiring human review are appended to wiki/_pending_questions.md. Each run logs token usage and estimated costs at the end.

Degraded runs. When a run exhausts its API call budget (see ATHENAEUM_MAX_API_CALLS above), it writes a wiki/_deferred_work.md manifest itemizing the raw files it did not process and ends with a warning-level Done (DEGRADED — budget exhausted) summary line — the machine-greppable signal that intake was deferred rather than completed. The deferred files stay on disk and are picked up automatically by the next run. The manifest is overwritten on every budget-tripped run and cleared by the next clean run (full, merge-only, or cluster-only). The cap is enforced at the entity-tier loop; merge-phase and re-resolve calls count toward the budget but do not themselves stop the run, so a merge-heavy run can overshoot the cap before enforcement kicks in. A degraded run still exits 0 by default; pass athenaeum run --strict-budget to make a budget-tripped run exit nonzero instead — opt-in, for exit-code-based alerting.

Data lifecycle & upgrade impact

Upgrade impact (0.10.0) — athenaeum run now MOVES and DELETES raw auto-memory by default. raw/auto-memory/ is an expiring intake queue, not a permanent store. As of 0.10.0, once the librarian has compiled a cluster into its canonical wiki/auto-<topic>.md entry and the contradiction detector has run clean, the move-then-retire pass (issue athenaeum#261) moves each non-contradictory raw fact into the wiki entry (as an origin-traced footnote) and git rms the raw file so it no longer re-enters the nightly loop. This is on by default. If you upgrade and run without reading this, your raw auto-memory files will start disappearing from the working tree — recoverable, but only from git history.

What is moved vs. held. Only non-contradictory clusters are retired. A cluster is held in the queue (never deleted) when the detector flags a contradiction, when the detection degraded (offline / API error / unparseable response), or when a member is referenced by an open entry in _pending_questions.md / _pending_merges.md. When in doubt, the pass keeps the raw.

Recovery is git-only. The pass refuses to run when knowledge_root is not a git repo, and it never hard-unlinks. Each retirement lands as two commits in your knowledge repo:

• Commit A — provenance snapshot. The raw intake about to be retired is committed first (scoped git add of exactly those files) so every deleted file is recoverable from history.
• Commit B — move + delete together. The wiki updates (new footnotes, retired: true marker) and the raw git rms land in a single commit, so the fact is never simultaneously absent from both the raw file and the wiki.

To recover a retired raw file, find commit B (or A) in your knowledge repo and git show/git checkout the path. Warning: because recovery depends entirely on git history, anything that rewrites or discards that history can lose retired raw permanently — git gc pruning unreachable objects, a squash/rebase that collapses the snapshot commits, or simply never committing (running on a dirty repo) / never pushing to a backup remote. If you rely on retired-raw recovery, keep the knowledge repo's history intact and pushed.

Pushing after every run (opt-in, issue athenaeum#284). A scheduled nightly run commits locally, but does not push by default — so origin silently drifts and the git-only recovery story only holds on the machine that ran the librarian. Two ways to turn on a post-run push so origin stays current:

athenaeum run --push               # one run: push after this run

# athenaeum.yaml — persistent opt-in
librarian:
  push_after_run: true
  # Optional; defaults are origin + the current branch's upstream.
  # push_remote: origin
  # push_branch: develop

The --push CLI flag overrides the yaml toggle. When enabled, athenaeum invokes git push (using the operator's ambient git auth — credential helper / SSH; no tokens or secrets are handled by athenaeum) after a successful run that produced at least one commit. --dry-run never pushes; a run with no new commits never pushes. A push failure is reported as a non-fatal warning (distinct log line athenaeum-push-failed:) — commits remain local and the next run retries (git push is idempotent).

--dry-run computes the exact same plan and logs a structured report without moving, deleting, or committing anything — use it to preview what a run would retire.

Disabling it. Move-then-retire stays on by default, but you can turn it off two ways:

athenaeum run --no-retire          # one run: skip the retire pass entirely

# athenaeum.yaml — persistent opt-out
librarian:
  retire: false

The --no-retire CLI flag overrides the yaml toggle. When disabled, raw auto-memory is neither moved into the wiki nor git rm'd; it stays in the intake queue and is re-examined on every run.

Related destructive operation. athenaeum auto-memory prune --apply is a second opt-in command that git rms pages (operational wiki/auto-*.md), with the same git-only recovery story as move-then-retire. It is dry-run by default — see Maintenance & inspection commands above.

Known limitations

Athenaeum is pre-1.0. These trade-offs are intentional for the current release line:

• No retrieval benchmarks yet. The hybrid-search claim rests on concrete failure modes (proper-noun collision, no-overlap semantic queries) and production use — not a published eval against mem0 / Letta / Zep / Cognee. If you need benchmarked recall@k on a closed corpus, pick a tool that publishes numbers. If you want a knowledge base that survives your tool choices, this is for you. PRs adding an eval harness are very welcome.
• FTS5 index rebuilds are non-atomic and unlocked. A shell hook and the librarian run rebuilding simultaneously can race; the window is small and single-user wikis do not hit it in practice, but hardened multi-writer safety remains future work. The athenaeum run / ingest / reindex (rebuild-index) commands are single-flight against each other via the run lock (issue athenaeum#309); the residual race is only with the example shell hook, which does not take that lock.
• The keyword search backend is a scan-on-query fallback. It reads every wiki page on every query; fine under ~1,000 entities, painful past that. Use search_backend: fts5 (default in the CLI and hooks) for any non-trivial wiki. The keyword backend exists as a zero-dependency baseline for tests and bootstrap.
• Tier 4 (human escalation) is a file, not a workflow. Conflicts land in wiki/_pending_questions.md; you read it and decide. No PR-opening, no Slack integration, no UI — on purpose, for now.
• Reasoning tiers exist but aren't wired into any default run path. See the note under "Key features" above — the tiered-reasoning pipeline is reachable only through one opt-in, default-off audit hook.
• Storage adapters are an extension point with one real in-tree consumer. See "Key features" above — is_embedded/is_recallable have no callers in src/ yet; only is_merge_eligible is wired.

Development

git clone https://github.com/Kromatic-Innovation/athenaeum.git
cd athenaeum
pip install -e ".[dev,vector]"   # matches CI; [dev] alone works if you won't touch search/clustering

pytest tests/ -v
ruff check src/ tests/

Branch flow

Athenaeum follows trunk-style development, with develop as the active branch and main as the released-revision pointer:

• develop is the active development branch and the GitHub default. All pull requests target develop.
• main carries the most recent released revision. Release tags (vX.Y.Z) live on main and trigger the PyPI release workflow.

Most users should install via pip install athenaeum (above). To work from source against the latest released revision instead of the active branch, clone and check out the latest tag:

git clone https://github.com/Kromatic-Innovation/athenaeum.git
cd athenaeum
git checkout "$(git describe --tags --abbrev=0)"

See CONTRIBUTING.md for the full promotion flow.

Where to go next

This README covers the shape of the system and how to run it. For depth on a specific piece:

• docs/why-athenaeum.md — full design rationale, comparison with mem0/Letta/Zep/Cognee/RAG/Claude memory, and production lessons.
• docs/recall-architecture.md — the hybrid FTS5 + vector recall path and its invariants.
• docs/contradiction-detection.md — contradiction pipeline, cross-scope modes, cost model.
• docs/conflict-resolution.md and docs/auto-resolve.md — the resolver's action taxonomy and auto-apply lane.
• docs/authority-manifest.md and docs/source-handles.md — source precedence and provenance handles.
• docs/security-posture.md — audience scoping, fail-closed enforcement, PII screening.
• docs/storage-adapter-contract.md — the storage extension point.
• docs/adapter-contract.md — writing a custom intake adapter.
• docs/configuration.md — every env var, yaml key, and CLI flag, with defaults and precedence.

Getting help

Rolling this out on a team? Open an issue or reach out via kromatic.com. We talk to teams working through agent-memory rollouts often and are happy to point at whatever's useful.

License

Apache 2.0 — see LICENSE.