szl-recall 0.3.6

Persistent memory layer for AI agents. Local-first. Inspectable. Framework-agnostic.

Recall

Persistent memory layer for AI agents. Local-first. Inspectable. Framework-agnostic.

PyPI: szl-recall  ·  GitHub: Sentient-Zero-Labs/agent-recall

Built as the implementation anchor for the Building Effective Tools for AI series by Sentient Zero Labs.

---

What Recall is

Recall is an MCP server that gives AI agents durable, structured memory across sessions.

An agent sends raw conversation text to Recall via store_memory. Recall extracts typed facts in the background using Claude Haiku — things like "user prefers Python for backend services" or "team decided to use FastAPI for the new service" — and stores them as searchable memories. When the agent needs context, it calls search_memories and gets back the most relevant memories ranked by a 4-component scoring model (relevance, recency, importance, access frequency).

Everything is local: one SQLite file, one process, no external services beyond the Anthropic API for extraction.

---

Capabilities

Memory types extracted from conversation text:

Type	Meaning	Example
preference	How the user likes to work	"User prefers dark mode and vim keybindings"
fact	Stated facts about the user or context	"User works at Acme Corp on the payments team"
decision	Choices made during conversation	"Team decided to use Postgres over MySQL for this project"
procedure	Processes the user described	"User's deploy process: test → staging → manual approval → prod"

Structured facts — preferences and facts are also stored with (entity, attribute, value) triples, enabling deterministic contradiction detection. When a new memory contradicts an existing one (same entity + attribute, different value), the old memory is automatically superseded. Superseded memories are excluded from all queries.

Hybrid search — queries are ranked using BM25Plus (keyword relevance) fused with optional dense vector similarity via RRF (Reciprocal Rank Fusion, k=60), then re-scored with a 4-component formula:

score = w_rrf · RRF(BM25, cosine)
      + w_recency · exp(−age_days / (1 + access_count))
      + w_import  · importance
      + w_strength · log(1 + access_count) / log(1 + max_access_count)

Weights sum to 1.0 and shift linearly with recency_weight param:
  recency_weight=0 → (w_rrf=0.70, w_recency=0.00, w_import=0.20, w_strength=0.10)
  recency_weight=1 → (w_rrf=0.40, w_recency=0.40, w_import=0.10, w_strength=0.10)

Two ingestion paths:

1. MCP tools — async-acknowledge: store_memory returns in <10ms, extraction runs in a background worker.
2. A2A protocol — synchronous task interface for agent-to-agent calls. Supports a consolidate_memories skill with a contradiction resolution flow (input-required state when conflicts are detected).

Security — tool descriptions are validated at startup against injection patterns (URLs, conditional behavior instructions, exfiltration patterns). Bearer tokens are SHA-256 hashed before storage. A 30-second hard timeout prevents runaway tool calls.

---

Install

pip install szl-recall

With optional dense vector search (requires ~500MB for the model on first run):

pip install "szl-recall[embeddings]"

---

Quick start

# 1. Set your Anthropic API key (used for background extraction)
export ANTHROPIC_API_KEY=sk-ant-...

# 2. Start the server — this initializes the database on first run
recall serve --port 8000

# In a new terminal:

# 3. Create an API token for your user
recall create-token my-agent --db recall.db

#   Output:
#     Token created for namespace: my-agent
#
#       Bearer <token>
#
#     Store this token securely — it will not be shown again.

Add to Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "recall": {
      "url": "http://localhost:8000/mcp",
      "transport": "streamable-http",
      "headers": {
        "Authorization": "Bearer <your-token>"
      }
    }
  }
}

Restart Claude Desktop. Recall's seven tools are now available in every conversation.

---

Seven MCP tools

store_memory

Stores conversation text for background extraction. Returns immediately (<10ms). Safe to retry with the same idempotency_key.

Arguments:
  text            str   — conversation transcript or fact to store
  topic           str   — context label (e.g. "engineering", "personal", "project-x")
  idempotency_key str   — (optional) unique key; duplicates are silently ignored. Auto-generated UUID if omitted.
  session_id      str   — (optional) conversation ID for provenance
  agent_id        str   — (optional) which agent submitted this

Returns:
  {status: "ok", data: {queued: true, job_id: "..."}}     — new job
  {status: "ok", data: {queued: false, cached: true}}     — duplicate key

What happens after: the extraction worker picks up the job, calls Claude Haiku with the text, extracts 0–5 typed memories, checks for contradictions with existing facts, and persists to SQLite. The job_id can be used to correlate log entries.

search_memories

Hybrid retrieval over all stored memories for the authenticated namespace.

Arguments:
  query          str     — natural language query
  limit          int     — max results (default 20, hard cap 50)
  recency_weight float   — 0.0–1.0 (default 0.3); higher surfaces newer memories
  mmr_lambda     float   — 0.0–1.0 (default 0.5); 0=max diversity, 1=pure relevance
  score_threshold float  — drop candidates below this hybrid score (default 0.0)
  max_tokens     int     — optional token budget; trims results to fit (4 chars/token)

Returns:
  {status: "ok", data: {results: [...memories], total: N}}

Each result includes: id, text, topic, type, importance, created_at.

MMR diversification: when sentence-transformers is installed, results are reranked using Max-Marginal Relevance so near-duplicate memories don't all crowd the top slots. Set mmr_lambda=0.3 for high diversity, mmr_lambda=1.0 to disable.

inspect_memories

Paginated list of all active memories (superseded memories are excluded).

Arguments:
  limit   int   — page size (default 20, max 50)
  offset  int   — pagination offset (default 0)

Returns:
  {status: "ok", data: {memories: [...], total: N, has_more: bool, next_offset: N|null}}

delete_memory

Permanently removes a memory. User-scoped — an agent can only delete its own memories.

Arguments:
  memory_id  str   — ID from inspect_memories or search_memories

Returns:
  {status: "ok", data: {deleted: "<id>"}}
  {status: "error", code: "MEMORY_NOT_FOUND"}

get_memory_stats

Fast health check. Returns counts of active memories by type and the number of pending extraction jobs.

Arguments: none

Returns:
  {
    status: "ok",
    data: {
      by_type: {preference: N, fact: N, decision: N, procedure: N},
      total: N,
      pending_extractions: N
    }
  }

consolidate_memories

Finds and merges semantically similar memories in a given topic using embedding-based clustering followed by an LLM merge pass. Reduces memory bloat and improves retrieval quality over time.

Arguments:
  topic                str    — topic to consolidate (e.g. "engineering")
  similarity_threshold float  — cosine similarity cutoff for clustering (default 0.85)
  dry_run              bool   — if true, returns a plan without modifying the database

Returns:
  {
    status: "ok",
    data: {
      groups_found: N,
      memories_consolidated: N,
      memories_created: N,
      deleted_ids: [...],
      created: [...],
      dry_run: bool
    }
  }
  {status: "error", code: "EMBEDDINGS_REQUIRED"}  — sentence-transformers not installed

What happens: active memories in the topic are embedded, clustered by cosine similarity, and each cluster is sent to Claude Haiku for a merge pass. Original memories are superseded (valid_until set); a single canonical memory is written in their place.

delete_namespace_data

Permanently deletes ALL data for the authenticated namespace. Irreversible. Requires exact confirmation string.

Arguments:
  confirm  str  — must be exactly "DELETE MY DATA" to proceed

Returns:
  {status: "ok", data: {tokens_revoked: N}}
  {status: "error", code: "CONFIRM_REQUIRED"}

What happens: deletes all memories, operations, A2A tasks, and tool call records for the namespace. Revokes all API tokens. Subsequent requests with any token for this namespace will return 401.

---

A2A protocol

Recall also implements the Agent-to-Agent (A2A) protocol for direct agent-to-agent calls. The agent card is published at GET /.well-known/agent-card.json.

Create a consolidation task:

POST /a2a
Authorization: Bearer <token>

{
  "skill": "consolidate_memories",
  "input": {
    "text": "Conversation transcript...",
    "topic": "engineering"
  }
}

→ 202 {id: "<task_id>", status: "submitted"}

Poll for completion:

GET /a2a/<task_id>

→ {id: "...", status: "working" | "completed" | "input-required", output: {...}}

Resolve contradictions (when status == "input-required"):

POST /a2a/<task_id>/resume

{
  "resolution": "keep_existing" | "keep_new" | "keep_both"
}

A2A vs MCP store_memory:

	MCP store_memory	A2A consolidate_memories
Latency	<10ms (async-acknowledge)	2–6s (synchronous extraction)
Contradiction handling	Auto-resolved via schema	input-required flow for ambiguous cases
Caller	Any MCP client	Another agent

---

Architecture

                      Agent / Claude Desktop
                             │
                    Bearer <token>
                             │
                             ▼
             ┌───────────────────────────────┐
             │       BearerAuthMiddleware     │  hash(token) → api_tokens
             │       TimeoutMiddleware        │  asyncio.wait_for(30s)
             │       LoggingMiddleware        │  ToolCallRecord per call
             └───────────────┬───────────────┘
                             │
               ┌─────────────┴──────────────┐
               │                            │
               ▼                            ▼
        POST /mcp                       POST /a2a
        FastMCP tools                   A2A task router
               │                            │
               │  store_memory              │  consolidate_memories
               │    ├── INSERT operations   │    ├── worker.extract() [sync]
               │    ├── enqueue job ──────► │    ├── detect contradictions
               │    └── return <10ms        │    └── bulk_store / input-required
               │                            │
               │  search_memories           GET /a2a/<id>   poll
               │    └── _hybrid_search      POST /a2a/<id>/resume
               │         ├── BM25Plus ranks
               │         ├── dense ranks (optional)
               │         ├── RRF fusion
               │         └── 4-component score
               │
               ▼
         ┌──────────────┐     asyncio.Queue
         │   SQLite DB  │ ◄──────────────────── ExtractionWorker
         │   (WAL mode) │                          │
         └──────────────┘                   Claude Haiku API
                                            extract → structured facts
                                            → _handle_contradiction
                                            → INSERT memories

Key invariants:

• Every DB operation is scoped to namespace — injected via ContextVar from auth middleware, never passed as a tool argument. A namespace can represent a user, agent, project, or any string identity.
• store_memory is idempotent: INSERT OR IGNORE + rowcount check makes concurrent retries safe.
• Active memories are always WHERE valid_until IS NULL. Superseded facts remain in the DB (for audit) but are excluded from all queries.
• On server restart, _recover_orphaned_operations() marks any stuck jobs as failed to prevent ghost jobs.

---

Database schema

Six tables in recall.db:

Table	Purpose
memories	Core store. Active rows have valid_until IS NULL.
operations	Idempotency + job lifecycle tracking for async extraction.
api_tokens	SHA-256 hashed bearer tokens per namespace.
tool_call_records	Structured audit log for every tool call (duration, tokens, cost).
schema_version	Migration tracking.
a2a_tasks	Persistent A2A task store. Survives server restart.

Key memories columns:

Column	Type	Description
text	TEXT	The memory content
type	TEXT	preference, fact, decision, procedure
entity	TEXT	Subject of structured fact (e.g. "user")
attribute	TEXT	Property (e.g. "preferred_language")
value	TEXT	Value (e.g. "Python")
valid_from	TEXT	When this fact became true
valid_until	TEXT	When superseded (NULL = still active)
importance	REAL	0.0–1.0, LLM-assigned
access_count	INT	Increments on retrieval (used in recency decay)
embedding	BLOB	L2-normalized float32 vector (BAAI/bge-small-en-v1.5)

---

Environment variables

Variable	Required	Default	Description
ANTHROPIC_API_KEY	Yes	—	Claude Haiku for background extraction
RECALL_DB_PATH	No	recall.db	Path to SQLite database file
RECALL_DB_URL	No	—	Postgres DSN (e.g. postgresql://user:pass@localhost/recall). If set, uses Postgres instead of SQLite.
RECALL_DECAY_LAMBDA	No	0.02	Decay rate — ~35-day half-life at default.
RECALL_DECAY_JOB_INTERVAL	No	3600	Seconds between decay runs.

---

Development

git clone https://github.com/Sentient-Zero-Labs/agent-recall
cd recall

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

# Run all tests (worker tests require ANTHROPIC_API_KEY)
pytest tests/ -v

# Fast tests only — no API key needed (~2s)
pytest tests/test_search.py tests/test_server.py tests/test_client.py tests/test_models.py -v

# With embeddings (requires sentence-transformers)
pip install -e ".[dev,embeddings]"
pytest tests/test_mmr.py -v   # MMR deduplication + diversity tests

# Live extraction tests (~8s)
ANTHROPIC_API_KEY=sk-ant-... pytest tests/test_worker.py -v -s

Test coverage by layer:

File	What it covers
test_client.py	SQLite layer — store, get, search, delete, pagination
test_models.py	MemoryUnit validation and serialization
test_server.py	HTTP layer — auth, MCP routing, response shapes, error codes
test_search.py	Search correctness (superseded filtering, idempotency) + ranking behavior
test_worker.py	Extraction pipeline — Haiku output structure, queue→DB path, stub fallback
test_mmr.py	MMR reranking, diversity, score_threshold, fallback
test_budget.py	Context budget trimming (max_tokens)
test_gdpr.py	delete_namespace_data — wrong confirm, full erasure, token revocation
test_postgres_backend.py	DB backend abstraction, placeholder translation, factory
test_decay.py	DecayWorker — decay scoring, access-count protection, run_once count
test_consolidation.py	consolidate_memories — clustering, LLM merge, dry-run mode

---

CLI reference

recall serve [--host 0.0.0.0] [--port 8000] [--db recall.db] [--reload]
recall create-token <namespace> [--db recall.db]
recall status [--db recall.db]

<namespace> can be any string: alice, agent:code-reviewer-v2, project:payments, etc.

Note: --db flag on recall serve now correctly overrides RECALL_DB_PATH env var.

---

Version history

• v0.3.4 (current): delete_namespace_data renamed from delete_user_data (consistent naming). store_memory idempotency_key is now optional — auto-generates a UUID when omitted.
• v0.3.3: user_id renamed to namespace across all DB columns, code, and CLI. DB migration runs automatically on startup. MMR bug fix: relevance signal now uses hybrid scores (not raw cosine) so mmr_lambda=1.0 correctly preserves full ranking order.
• v0.3.0–v0.3.2: MMR diversification (mmr_lambda), context budget trimming (max_tokens), GDPR erasure (delete_namespace_data, 7th tool), A2A task persistence (survives restarts), Postgres backend abstraction (RECALL_DB_URL), --db CLI bug fix, TROUBLESHOOTING.md, sentence-transformers 5.5.0 validated.
• v0.2: Hybrid BM25Plus+RRF search, 4-component scoring, structured fact extraction (entity/attribute/value), deterministic contradiction detection, A2A consolidation worker.
• v0.1: SQLite + BM25 search, 5 MCP tools, async extraction via Claude Haiku, bearer auth.

Roadmap:

• Multi-namespace admin API (list/delete namespaces)
• Webhook notifications on extraction completion
• owner_id + agent_id composite isolation for agent-per-user memory separation (v0.4)

---

License

MIT — see LICENSE.

---

Built by Sentient Zero Labs. Newsletter: read.sentientzerolabs.com.