agent-memory

English | 简体中文

Zero-config, traceable, MCP-native long-term memory for agents.

agent-memory targets a gap in the current memory stack: a local-first engine that works with pip install, runs on pure SQLite, and makes memory evolution explainable instead of opaque.

Install from PyPI with pip install agent-memory-engine.

Current packaged release: 0.2.1.

Documentation

• English docs index: docs/README.md
• 中文文档索引: docs/zh-CN/README.md
• Delivery tutorial: docs/project-delivery-and-tutorial.md
• MCP guide: docs/mcp-integration.md
• Release guide: docs/release-and-pypi.md
• Benchmark report: docs/benchmark-results.md

Why this exists

• Mem0 proves demand, but pulls in heavier infra such as Neo4j or Qdrant.
• Local agents and personal copilots need a memory layer that is easy to embed, debug, export, and ship.
• The project spans a broad technical surface area: storage, retrieval, ranking, decay, provenance, conflict handling, MCP, and evaluation.

Current Status

• SQLite backend with WAL, FTS5, audit log, evolution log, entity index, and causal parent links
• Go service workspace with SQLite storage engine, schema migration, REST gateway, gRPC server, auth hooks, metrics, tracing bootstrap, and Cobra CLI
• Schema indexes for type, layer, recency, trust, source, relation, and audit hot paths
• Python SDK via MemoryClient
• Python MemoryClient now supports embedded and remote modes through SQLiteBackend and RemoteBackend
• In service mode, fused retrieval orchestration can run inside the Go service over REST/gRPC
• Rule-based intent router with Reciprocal Rank Fusion
• Adaptive forgetting utilities with dual-threshold layer transitions
• Heuristic conflict detection with contradiction edges and trust-score adjustment
• Optional LLM-backed conflict adjudication for top semantic candidates
• Governance helpers for health reports, audit reads, and JSONL export/import
• Optional MCP server and REST API adapters with dependency-friendly fallbacks
• sqlite-vec integration with safe fallback to Python cosine scan when unavailable
• Deterministic local fallback embeddings for testability and zero-friction startup
• LLM-first conversation extraction with heuristic fallback
• Trace graph reports with ancestors, descendants, relations, and evolution history
• Idempotent maintenance cycle for decay, promotion/demotion, conflict upkeep, and consolidation
• Benchmark helpers and LOCOMO-Lite style starter data

Quickstart

pip install agent-memory-engine

agent-memory store "User prefers SQLite for local-first agents." --source-id demo
agent-memory search "Why SQLite?"
agent-memory health

For development:

pip install -e '.[dev]'
.venv/bin/python -m pytest -q

from agent_memory import MemoryClient

client = MemoryClient()
item = client.add(
    "The user prefers SQLite for local-first agent projects.",
    source_id="demo-session",
)

results = client.search("What database does the user prefer?")
print(results[0].item.content)

trace = client.trace_graph(item.id)
print(trace.descendants)

health = client.health()
print(health.suggestions)

Service mode

make proto
cd go-server && go run ./cmd/server

export AGENT_MEMORY_MODE=remote
export AGENT_MEMORY_GO_SERVER_URL=http://127.0.0.1:8080
export AGENT_MEMORY_GRPC_TARGET=127.0.0.1:9090
agent-memory search "Why SQLite?"

Architecture

graph TD
    A["Python SDK / MCP"] --> B["MemoryClient"]
    B --> C{"Mode"}
    C -->|"embedded"| D["SQLiteBackend (Python)"]
    C -->|"remote"| E["RemoteBackend"]
    E --> F["Go REST / gRPC"]
    F --> G["SQLite Storage Engine"]
    G --> H[("SQLite + WAL + vector fallback")]
    B --> I["Intent Router / Conflict / Trust"]

Core components

• src/agent_memory/client.py — high-level SDK entry point
• src/agent_memory/storage/remote_backend.py — REST/gRPC bridge to the Go service
• proto/memory/v1/ — shared Protobuf contracts
• go-server/cmd/server/main.go — Go service entrypoint with graceful shutdown
• go-server/internal/storage/sqlite.go — Go storage engine
• go-server/internal/gateway/handler.go — Go REST handlers
• go-server/internal/grpc/server.go — Go gRPC implementation
• src/agent_memory/storage/sqlite_backend.py — SQLite persistence, FTS, vector fallback, trace queries
• src/agent_memory/controller/router.py — intent-aware retrieval routing and RRF fusion
• src/agent_memory/controller/forgetting.py — Ebbinghaus-inspired adaptive forgetting
• src/agent_memory/controller/conflict.py — contradiction detection and conflict records
• src/agent_memory/controller/consolidation.py — overlap grouping and merge-draft generation
• src/agent_memory/controller/trust.py — multi-factor trust scoring
• src/agent_memory/governance/health.py — stale/orphan/conflict monitoring
• src/agent_memory/interfaces/mcp_server.py — eight MCP tools
• src/agent_memory/extraction/pipeline.py — conversation-to-memory extraction
• benchmarks/ — storage/retrieval microbenchmarks and synthetic eval seeds

Design choices

• SQLite + WAL keeps deployment zero-config while fitting agent workloads: many reads, occasional writes.
• Rule routing over LLM routing keeps routing latency predictable and testable.
• RRF instead of score averaging avoids calibration problems across lexical, entity, and semantic retrieval.
• sqlite-vec plus fallback gives C/SQL vector search when available while keeping the package runnable everywhere.
• Soft delete preserves provenance and causal trace integrity.
• Hash fallback embeddings make the package runnable even before a local embedding model is available.
• Unique relation edges keep maintenance idempotent and health metrics stable.

Project layout

agent-memory/
├── deploy/
├── go-server/
├── proto/
├── docs/plans/
├── examples/
├── src/agent_memory/
│   ├── controller/
│   ├── embedding/
│   ├── extraction/
│   └── storage/
└── tests/

Benchmarks

Synthetic LOCOMO-Lite run on the bundled starter dataset (30 dialogues / 150 questions):

Metric	agent-memory	Semantic-only baseline
Overall hit rate	50.0%	23.3%
Factual recall	53.3%	6.7%
Temporal recall	36.7%	3.3%
Causal recall	53.3%	6.7%
p95 retrieval latency	16.64ms	11.50ms

• Full report: docs/benchmark-results.md
• Re-run locally: python benchmarks/locomo_lite/evaluate.py

MCP Usage

Install MCP support and launch the stdio server:

pip install -e .[mcp]
python -m agent_memory.interfaces.mcp_server

Claude Desktop configuration:

{
  "mcpServers": {
    "agent-memory": {
      "command": "python",
      "args": ["-m", "agent_memory.interfaces.mcp_server"],
      "env": {
        "AGENT_MEMORY_DB_PATH": "/absolute/path/to/default.db"
      }
    }
  }
}

Typical tools:

• memory_store — store a memory with provenance
• memory_search — run intent-aware retrieval
• memory_trace — inspect causal ancestry and evolution
• memory_health — inspect stale/conflict/orphan metrics

More details: docs/mcp-integration.md

Demos

• python examples/demo_cross_session.py --db /tmp/agent-memory-demo.db
• python examples/interactive_chat.py --db chat_memory.db --provider none
• python examples/mcp_server.py

Release Notes

• Changelog: CHANGELOG.md
• benchmarks/locomo_lite/latest_results.json is regenerated by the evaluation script
• docs/screenshots/ is reserved for verified MCP client screenshots
• Delivery record and full tutorial: docs/project-delivery-and-tutorial.md
• Release and PyPI guide: docs/release-and-pypi.md
• Expansion and optimization review: docs/plans/2026-03-24-agent-memory-expansion-review.md

Dev Notes

• Run all tests with .venv/bin/python -m pytest -q
• Use the built-in CLI with agent-memory --help
• sqlite-vec is installed as a package dependency; if the extension cannot be loaded at runtime, vector search safely falls back to Python cosine scan
• Try microbenchmarks with python benchmarks/bench_storage.py and python benchmarks/bench_retrieval.py
• Try the demo runner with python examples/benchmark_runner.py