ai-knot 0.6.0

Agent Knowledge Layer. Extract. Store. Retrieve. Any backend.

ai-knot

Agent knowledge layer — distills conversations into structured facts, retrieves what matters, forgets the rest.

Most agent frameworks treat memory as a log. ai-knot treats it as a knowledge base. It extracts facts from conversations, scores them by importance, and retrieves only what's relevant when building the next prompt. Pluggable storage, six LLM providers, no vendor lock-in.

---

The problem

Most frameworks store everything — messages, tool calls, system prompts, the whole log. That's fine until you're paying to inject six months of conversation history into every request, most of which has nothing to do with what the user asked.

The log grows to 400k tokens. The model needs maybe 300 of those for the next turn — but there's no obvious way to know which ones without reading all of them first. ai-knot solves this by keeping a distilled knowledge base instead of a raw log.

1000 messages  (~400k tokens)
    ↓ LLM extraction + ATC verification
~12 facts      (~300 tokens)
    ↓ BM25 retrieval
3–5 facts injected into the next prompt

---

Install

Python:

pip install ai-knot

# With OpenAI for LLM extraction:
pip install "ai-knot[openai]"

# With PostgreSQL backend:
pip install "ai-knot[postgres]"

# With MCP server (Claude Desktop / Claude Code):
pip install "ai-knot[mcp]"

Node.js / TypeScript (requires Python 3.11+ in PATH):

npm install ai-knot

---

Quickstart (30 seconds)

from ai_knot import KnowledgeBase

kb = KnowledgeBase(agent_id="my_agent")

# Add facts manually
kb.add("User is a senior backend developer at Acme Corp",
       type="semantic", importance=0.95)
kb.add("User prefers Python, dislikes async code",
       type="procedural", importance=0.85)

# Or extract automatically from a conversation
from ai_knot import ConversationTurn
turns = [
    ConversationTurn(role="user",      content="I deploy everything in Docker"),
    ConversationTurn(role="assistant", content="Got it, I'll use Docker examples"),
]
kb.learn(turns, provider="openai", api_key="sk-...")  # LLM extracts + stores relevant facts

# At inference time — get what matters
context = kb.recall("how should I write this deployment script?")
# -> "[procedural] User prefers Python, dislikes async code
#     [semantic]   User deploys everything in Docker
#     [semantic]   User is a senior backend developer at Acme Corp"

# Inject into your prompt
response = openai_client.chat(...,
    system=f"You are a helpful assistant.\n\n{context}")

---

Performance

Benchmarks run on Ubuntu (ubuntu-latest, GitHub Actions). Full benchmark history →

Retrieval latency (BM25, in-process)

Measured with pytest-benchmark, pedantic mode, 20 rounds after 3 warmups. BM25Retriever.search() is O(n) — IDF and BM25 scores are recomputed on every call.

Facts in memory	p50	p95	QPS
100	~1 ms	~3 ms	~800
1 000	~8 ms	~25 ms	~100
10 000	~80 ms	~200 ms	~12

Numbers are indicative. Run pytest tests/test_performance.py -m slow --benchmark-only locally for hardware-accurate results.

Full-stack recall latency (storage I/O + decay + BM25)

KnowledgeBase.recall() reads storage on every call. YAML adds ~10–50 ms I/O overhead; SQLite is lower-variance at scale.

Backend	Facts	p50	p95
YAML	1 000	~30 ms	~80 ms
SQLite	1 000	~20 ms	~50 ms

MCP tool call round-trip (stdio transport)

Measured end-to-end: Python subprocess spawn is one-time; per-call overhead is JSON serialization + tool execution.

Tool	Facts	p50	p95
add	—	~15 ms	~80 ms
recall	50	~20 ms	~100 ms
stats	—	~5 ms	~20 ms

Context: pure MCP stdio JSON-RPC overhead is ~10 ms P95 with no tool execution (tmdevlab MCP benchmark). Anthropic recommends keeping agent memory tool latency under 100 ms (Reduce Latency docs). Use storage="sqlite" for lower variance at scale.

---

What ai-knot keeps — and what it drops

Pass it a conversation and it calls your LLM to figure out what's worth keeping — preferences, recurring patterns, explicit facts. Greetings, clarifications, filler — those don't make the cut.

What happened in the conversation:         What ai-knot stores:
---                                        ---
"hey"                                      X skipped
"thanks"                                   X skipped
"ok got it"                                X skipped
"I really hate working with async"         -> "User dislikes async code"
"by the way we deploy on kubernetes"       -> "User deploys on Kubernetes"
"can you make it shorter please"           -> "User prefers concise responses"

Signal, not noise. Importance scores, power-law retention decay, ATC verification, deduplication — built in.

---

Conflict resolution — no more duplicate facts

learn() cross-checks new facts against everything already stored before inserting anything. If a new fact is semantically similar to an existing one (Jaccard similarity ≥ 0.7 by default), the existing fact is reinforced instead of a duplicate being created:

kb.add("User deploys on Docker")
kb.add("User deploys with Docker Compose")  # similar enough -> reinforced, not duplicated

# Control the threshold per call:
kb.learn(turns, provider="openai", conflict_threshold=0.8)

Importance is bumped by 0.05 (capped at 1.0) each time reinforcement fires — the knowledge base naturally weights well-confirmed facts higher over time.

---

Snapshots — version your knowledge base

Save and restore the complete state of a knowledge base at any point in time:

kb.add("User prefers Python")
kb.add("User deploys on Docker")

kb.snapshot("before_refactor")      # save current state

kb.add("User switched to Go")       # state changes
kb.forget(some_fact_id)

kb.restore("before_refactor")       # atomically roll back

# List all saved snapshots:
names = kb.list_snapshots()         # ["before_refactor"]

# Compare two snapshots (or current state):
diff = kb.diff("before_refactor", "current")
print(f"Added: {[f.content for f in diff.added]}")
print(f"Removed: {[f.content for f in diff.removed]}")

Both YAML and SQLite backends support snapshots. YAML stores them under .ai_knot/{agent_id}/snapshots/. SQLite stores them in the same database file.

---

MCP server — use ai-knot from Claude Desktop or Claude Code

ai-knot ships a native MCP server. Install it and Claude can call add, recall, forget, and snapshot as tools — without any Python code on your end:

pip install "ai-knot[mcp]"

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "ai-knot": {
      "command": "ai-knot-mcp",
      "env": {
        "AI_KNOT_AGENT_ID": "myagent",
        "AI_KNOT_STORAGE": "sqlite",
        "AI_KNOT_DB_PATH": "/absolute/path/to/memory.db"
      }
    }
  }
}

Available tools: add, recall, recall_json, forget, list_facts, stats, snapshot, restore, list_snapshots.

TypeScript agents: always use recall_json — it returns a stable JSON array ([] when empty). recall returns a plain string and "No relevant facts found." on empty — harder to parse reliably.

Environment variables:

Variable	Default	Description
AI_KNOT_AGENT_ID	default	Agent namespace
AI_KNOT_STORAGE	sqlite	sqlite (recommended) or yaml
AI_KNOT_DATA_DIR	.ai_knot	Base dir for file backends (use absolute path)
AI_KNOT_DB_PATH	—	Full path to SQLite file (overrides DATA_DIR for sqlite)

Note: Claude Desktop launches processes from a non-interactive shell where cwd is undefined. Always set AI_KNOT_DATA_DIR or AI_KNOT_DB_PATH to an absolute path.

---

OpenClaw integration

Which path should I use?

Situation	Solution
OpenClaw TypeScript app (recommended)	generate_mcp_config() → paste into ~/.openclaw/openclaw.json
Python agent (LangChain, LangGraph, CrewAI)	OpenClawMemoryAdapter(kb)

ai-knot works as an OpenClaw memory backend via MCP. Two steps:

pip install "ai-knot[mcp]"   # installs the ai-knot-mcp entry point

Note: ai-knot (without [mcp]) does not install ai-knot-mcp. The config will be generated but OpenClaw won't find the command.

Generate the config snippet:

import json
from ai_knot.integrations.openclaw import generate_mcp_config

print(json.dumps(generate_mcp_config("my_agent"), indent=2))

Paste the output into your OpenClaw config file:

• macOS / Linux: ~/.openclaw/openclaw.json
• Windows: %APPDATA%\OpenClaw\openclaw.json

Your agent will have access to all ai-knot tools: add, recall, recall_json, forget, list_facts, list_snapshots, stats, snapshot, restore.

For Python-native agents (LangChain, LangGraph, CrewAI), use the adapter class instead:

from ai_knot import KnowledgeBase
from ai_knot.integrations.openclaw import OpenClawMemoryAdapter

kb = KnowledgeBase("my_agent")
memory = OpenClawMemoryAdapter(kb)

memory.add([{"role": "user", "content": "Deploy on Fridays"}])
results = memory.search("deployment schedule")
memory.update(results[0]["id"], "Deploy on Thursdays")
memory.delete(results[0]["id"])

Multi-turn extraction: add() stores only the last user message and emits a warning if the list has more than one user message. For extracting multiple facts from a full conversation, use kb.learn(turns, api_key=...) directly.

update() assigns a new ID. The old fact is deleted and a new one is created. If you need to hold a stable reference, call delete() + add() yourself and record the returned ID.

---

Pluggable storage backends

No vendor lock-in. Swap backends with one line:

from ai_knot import KnowledgeBase
from ai_knot.storage import YAMLStorage, SQLiteStorage

# Development — zero infra:
kb = KnowledgeBase(agent_id="bot", storage=YAMLStorage())

# Production — single server:
kb = KnowledgeBase(agent_id="bot",
    storage=SQLiteStorage(db_path="/data/agent.db"))

Same API. Same code. Different storage.

---

Initialization — storage + LLM provider together

Storage is set once on KnowledgeBase. The LLM provider can be set at init (recommended for production) or passed per learn() call. They are independent and combine freely:

from ai_knot import KnowledgeBase, ConversationTurn
from ai_knot.storage import SQLiteStorage

# Option A: configure provider once at init (recommended)
kb = KnowledgeBase(
    agent_id="assistant",
    storage=SQLiteStorage(db_path="./agent.db"),
    provider="openai",
    api_key="sk-...",          # or reads OPENAI_API_KEY from env if omitted
)
kb.learn(turns)                # no credentials needed per call
kb.learn(more_turns)           # same provider reused

# Option B: pass provider per call (legacy, still supported)
kb = KnowledgeBase(agent_id="assistant", storage=SQLiteStorage(db_path="./agent.db"))
turns = [
    ConversationTurn(role="user", content="I deploy everything in Docker"),
    ConversationTurn(role="assistant", content="Got it!"),
]
kb.learn(turns, provider="openai")           # reads OPENAI_API_KEY from env
kb.learn(turns, provider="openai",    api_key="sk-...")      # explicit key
kb.learn(turns, provider="anthropic", api_key="sk-ant-...")  # Claude
kb.learn(turns, provider="openai-compat",                    # any compatible API
         api_key="...", base_url="http://localhost:8000/v1")

# Recall never calls the LLM — no provider needed
context = kb.recall("how should I deploy this?")

Mix and match: any storage backend with any LLM provider.

---

Retrieval with relevance scores

recall_facts_with_scores() returns each fact together with its numeric relevance score. The score is a hybrid value combining BM25 similarity to the query, retention (power-law decay), and the fact's importance — higher is more relevant.

Use it when you need to filter or rank facts programmatically rather than inject them directly into a prompt:

scored = kb.recall_facts_with_scores("Docker deployment", top_k=5)
for fact, score in scored:
    print(f"[{score:.2f}] [{fact.type.value}] {fact.content}")
# [0.87] [procedural] User deploys everything in Docker
# [0.61] [episodic] Deploy failed last Tuesday at 3 PM

# Keep only highly relevant facts
relevant = [fact for fact, score in scored if score >= 0.5]

vs recall_facts() — use recall_facts() when you just need the Fact objects; use recall_facts_with_scores() when scores matter for downstream logic.

---

Memory types

Type	When to use	Example
semantic	Stable facts about the user or world	"User works at Sber", "Stack is Python + FastAPI"
procedural	How the user wants things done	"Always use type hints", "Prefer pytest over unittest"
episodic	Specific past events with time context	"Deploy failed last Tuesday at 3 PM", "User approved the v2 design on Monday"

Not sure which type to use? Default semantic covers most cases. Use procedural for preferences and rules; episodic for dated events you might want to forget sooner.

---

Forgetting (why it matters)

Accumulating everything makes agents worse, not better. Irrelevant facts pollute the context window — this is called context rot.

ai-knot uses a power-law decay curve (Wixted & Ebbesen, 1997) — empirically superior to exponential decay (R²=98.9% vs 96.3%):

retention(t) = (1 + t / (9 × stability))^(-1)

stability = 336h × importance × (1 + ln(1 + access_count))
-> high importance + frequently accessed = remembered for months
-> low importance + never accessed     = forgotten in weeks

Power-law has a heavier tail than exponential — important facts persist realistically over months instead of vanishing after days. Facts accessed often get reinforced. Stale facts fade automatically.

Do you need to call decay() manually? No — decay is applied automatically inside every recall() call. For facts that are never recalled (e.g. background knowledge your agent doesn't actively query), run kb.decay() in a daily cron job to keep retention scores current.

---

CLI

ai-knot show   my_agent            # list all stored facts
ai-knot recall my_agent "query"    # test retrieval
ai-knot add    my_agent "fact"     # add a fact
ai-knot stats  my_agent            # counts, avg importance, retention
ai-knot decay  my_agent            # apply forgetting curve
ai-knot clear  my_agent            # wipe knowledge base
ai-knot export my_agent out.yaml   # backup to file
ai-knot import my_agent in.yaml    # restore from backup

---

How knowledge looks on disk (YAML backend)

# .ai_knot/my_agent/knowledge.yaml — readable, editable, Git-trackable

a1b2c3:
  content: "User is a senior backend developer at Acme Corp"
  type: semantic
  importance: 0.95
  retention_score: 0.91
  access_count: 12
  created_at: '2026-03-01T10:00:00+00:00'
  last_accessed: '2026-03-27T09:00:00+00:00'
  tags: [user_profile, work]

d4e5f6:
  content: "User prefers Python, dislikes async patterns"
  type: procedural
  importance: 0.85
  retention_score: 0.88
  access_count: 34

Edit it by hand. Commit it to Git. Roll back when needed.

---

Batch fact insertion — add_many()

Insert multiple pre-extracted facts in a single storage round-trip, without any LLM call:

# Plain strings — use method-level defaults for type/importance/tags
kb.add_many(["User deploys on Fridays", "User uses Docker", "Stack: Python + FastAPI"])

# Dicts for full control per fact
kb.add_many([
    {"content": "User is a senior backend engineer", "type": "semantic", "importance": 0.95},
    {"content": "Always use type hints", "type": "procedural", "importance": 0.8},
    {"content": "Sprint demo went well", "type": "episodic", "importance": 0.6},
])

# Mix strings and dicts — strings use method defaults
kb.add_many(
    ["Quick fact"],
    type=MemoryType.PROCEDURAL,
    importance=0.7,
)

Useful when facts come from an external source, are pre-processed by another tool, or the LLM extraction step is handled upstream.

---

Async API

All blocking operations have async variants that run in a thread-pool executor, keeping the asyncio event loop free during LLM HTTP calls:

Sync	Async
kb.learn(turns, ...)	await kb.alearn(turns, ...)
kb.recall(query)	await kb.arecall(query)
kb.recall_facts(query)	await kb.arecall_facts(query)

import asyncio
from ai_knot import KnowledgeBase, ConversationTurn

kb = KnowledgeBase(agent_id="bot", provider="openai", api_key="sk-...")

# FastAPI handler — never blocks the event loop
async def handle_message(turns: list[ConversationTurn]) -> str:
    await kb.alearn(turns)
    return await kb.arecall("current topic")

# Concurrent extraction for multiple agents
kb_a = KnowledgeBase(agent_id="a", provider="openai", api_key="sk-...")
kb_b = KnowledgeBase(agent_id="b", provider="openai", api_key="sk-...")
results = await asyncio.gather(
    kb_a.alearn(turns_a),
    kb_b.alearn(turns_b),
)

---

learn() options: timeout and batch_size

# Abort slow LLM calls after 10 seconds (default: 30 s)
kb.learn(turns, provider="openai", api_key="sk-...", timeout=10.0)

# Split long conversations into chunks of 10 turns per LLM call (default: 20)
# Prevents silent fact loss when the LLM truncates a large JSON response
kb.learn(turns, provider="openai", api_key="sk-...", batch_size=10)

---

LLM providers

ai-knot ships with 6 providers for fact extraction:

Provider	Name	Env var
OpenAI	openai	OPENAI_API_KEY
Anthropic (Claude)	anthropic	ANTHROPIC_API_KEY
GigaChat	gigachat	GIGACHAT_API_KEY
Yandex GPT	yandex	YANDEX_API_KEY
Qwen	qwen	QWEN_API_KEY
Any OpenAI-compatible	openai-compat	LLM_API_KEY

kb.learn(turns, provider="anthropic")  # uses ANTHROPIC_API_KEY from env
kb.learn(turns, provider="gigachat", api_key="...")
kb.learn(turns, provider="openai-compat", api_key="...", base_url="http://localhost:8080/v1")

---

OpenAI integration

from ai_knot import KnowledgeBase
from ai_knot.integrations.openai import MemoryEnabledOpenAI

kb = KnowledgeBase(agent_id="assistant")
kb.add("User prefers Python")
kb.add("User deploys on Docker")

client = MemoryEnabledOpenAI(knowledge_base=kb)

import openai
openai_client = openai.OpenAI()

messages = [{"role": "user", "content": "Write me a deployment script"}]
enriched = client.enrich_messages(messages)

response = openai_client.chat.completions.create(
    model="gpt-4o",
    messages=enriched,
)

---

Architecture

Conversation Turns
       |
[ Extractor ]         LLM-based distillation -> structured facts
       |                 + ATC verification (Broder, 1997)
[ KnowledgeBase ]     importance scoring + deduplication + power-law decay
       |
[ Storage Adapter ]   YAML / SQLite / PostgreSQL (Mongo, Qdrant planned)
       |
[ BM25 Retriever ]    Okapi BM25 (zero deps) + Embeddings (planned)
       |
Context String        injected into agent system prompt

Why BM25 instead of embeddings? Embeddings need either an API call or a 500 MB local model just to recall which language the user prefers. For knowledge bases up to ~10k facts, BM25 with hybrid scoring (term relevance + retention + importance) is fast, deterministic, and requires zero extra setup. BM25 (Robertson & Zaragoza, 2009) handles term saturation and document length normalization better than raw TF-IDF. Semantic embeddings are on the roadmap for larger knowledge bases where keyword overlap isn't reliable.

ATC verification: Every LLM-extracted fact is verified against the source text using Asymmetric Token Containment (ATC). Facts where fewer than 60% of tokens appear in the source are flagged as supported=False, preventing hallucinated facts from polluting the knowledge base.

Known limitation: extraction quality depends on the LLM. GPT-4o extracts nuanced procedural facts reliably; smaller models (gpt-3.5-turbo, haiku) occasionally miss implicit preferences or conflate episodic events with semantic facts. When accuracy matters, use a capable model for learn().

---

Examples

1. Manual add + recall (no LLM required)

from ai_knot import KnowledgeBase, MemoryType

kb = KnowledgeBase(agent_id="assistant")
kb.add("User prefers Python",          type=MemoryType.PROCEDURAL, importance=0.9)
kb.add("User deploys with Docker",     importance=0.85)
kb.add("Deploy failed last Tuesday",   type=MemoryType.EPISODIC,   importance=0.4)

context = kb.recall("how to deploy?")
# -> "[procedural] User prefers Python
#     [semantic]   User deploys with Docker"

2. SQLite + OpenAI

from ai_knot import KnowledgeBase, ConversationTurn
from ai_knot.storage import SQLiteStorage

kb = KnowledgeBase(agent_id="bot", storage=SQLiteStorage(db_path="./bot.db"))
turns = [ConversationTurn(role="user", content="I work with Python and FastAPI")]
kb.learn(turns, provider="openai")          # reads OPENAI_API_KEY from env
context = kb.recall("what stack does user use?")

3. YAML storage + Anthropic (Claude)

from ai_knot import KnowledgeBase, ConversationTurn
from ai_knot.storage import YAMLStorage

kb = KnowledgeBase(agent_id="bot", storage=YAMLStorage(base_dir=".ai_knot"))
turns = [ConversationTurn(role="user", content="Always write tests with pytest")]
kb.learn(turns, provider="anthropic", api_key="sk-ant-...")
# Facts are saved to .ai_knot/bot/knowledge.yaml — readable, Git-trackable

4. PostgreSQL + any OpenAI-compatible endpoint

from ai_knot import KnowledgeBase, ConversationTurn
from ai_knot.storage import create_storage

storage = create_storage("postgres", dsn="postgresql://user:pass@db:5432/ai-knot")
kb = KnowledgeBase(agent_id="assistant", storage=storage)
turns = [ConversationTurn(role="user", content="Prefer concise answers")]
kb.learn(turns, provider="openai-compat",
         api_key="...", base_url="http://localhost:8000/v1")

5. Per-customer knowledge (support agent)

from ai_knot import KnowledgeBase

def handle_ticket(customer_id: str, message: str) -> str:
    kb = KnowledgeBase(agent_id=f"customer_{customer_id}")
    context = kb.recall(message)
    # Agent sees: past issues, preferences, tier — specific to this customer
    return context

6. Coding agent with project context

from ai_knot import KnowledgeBase, MemoryType
from ai_knot.storage import YAMLStorage

kb = KnowledgeBase(agent_id="project", storage=YAMLStorage(".ai_knot"))
kb.add("Stack: FastAPI + PostgreSQL + Docker",  importance=1.0)
kb.add("No unittest — use pytest only",         type=MemoryType.PROCEDURAL, importance=0.9)
kb.add("All endpoints require JWT auth",        importance=0.95)
# Commit .ai_knot/ to Git — new team members clone the context

7. Shared knowledge across multiple agents

from ai_knot import KnowledgeBase
from ai_knot.storage import SQLiteStorage

storage = SQLiteStorage(db_path="./team.db")
researcher = KnowledgeBase(agent_id="team_alpha", storage=storage)
writer     = KnowledgeBase(agent_id="team_alpha", storage=storage)

researcher.add("API rate limit is 100 req/s")
context = writer.recall("rate limits")  # sees researcher's facts instantly

8. Stats and forgetting curve

from ai_knot import KnowledgeBase

kb = KnowledgeBase(agent_id="assistant")
kb.add("User likes dark mode")
kb.add("User timezone is UTC+3")

stats = kb.stats()
print(f"Facts: {stats['total_facts']}")
print(f"Avg importance: {stats['avg_importance']:.2f}")
print(f"By type: {stats['by_type']}")

kb.decay()  # apply power-law forgetting curve — stale facts lose retention score

---

Roadmap

• Core KnowledgeBase (extraction + storage + retrieval)
• Power-law forgetting curve (Wixted & Ebbesen, 1997)
• YAML + SQLite backends
• OpenAI integration
• CLI
• PostgreSQL backend
• Conflict resolution in learn() (cross-session deduplication)
• Snapshots (snapshot, restore, diff)
• MCP server (Claude Desktop / Claude Code)
• npm package (npm install ai-knot)
• OpenClaw integration (OpenClawMemoryAdapter + generate_mcp_config)
• Scored retrieval (recall_facts_with_scores)
• BM25 retrieval with p95-clip normalization (replaces TF-IDF)
• ATC fact verification guardrail (Broder, 1997)
• Offline eval framework (P@k, R@k, MRR, nDCG, bootstrap CI)
• CI quality gates (eval-smoke, eval-full)
• MongoDB backend
• Qdrant + Weaviate backends
• Semantic embeddings (sentence-transformers / OpenAI)
• LangChain / CrewAI integrations
• Web UI knowledge inspector
• REST API / sidecar mode

---

Why not just use Mem0 / Zep / LangMem?

	ai-knot	Mem0	Zep	LangMem
Self-hosted	Yes	Partial	Yes	Yes
No cloud required	Yes	No	No	Yes
Pluggable storage	Yes	No	No	No
Human-readable store	Yes	No	No	No
Setup time	30 sec	10 min	30 min	5 min
Framework-agnostic	Yes	Partial	Partial	LangGraph only
Forgetting curve (power-law)	Yes	No	No	No
Fact verification (ATC)	Yes	No	No	No
Offline eval framework	Yes	No	No	No
Snapshots + diff	Yes	No	No	No
MCP server	Yes	No	No	No
Free forever	Yes (MIT)	No	No	Yes

---

Contributing

PRs welcome. Especially looking for: storage backend implementations, integration adapters, retrieval strategies.

---

License

MIT

---

Found a bug or a missing backend? Open an issue. Built something with it? We'd like to hear.