memry 0.2.0

The open, self-hostable memory layer for AI agents. MCP-native, local-first, research-grade.

Memry

memry.tech - the open, self-hostable memory layer for AI agents. One pip install, zero services, plugged into any agent over MCP - with the intelligence layer (extraction, reconciliation, temporal invalidation, decay, context construction) as first-class, replaceable research code.

pip install memry
memry mcp          # ← your agent now has long-term memory

• MCP-native - works with Claude Code, Claude Desktop, Cursor, Windsurf, Codex, and any other MCP client, over stdio or streamable HTTP.
• Local-first - a single SQLite file. No vector DB, no Postgres, no cloud. Works with zero API keys (FTS5 BM25 + deterministic hash embeddings), gets smarter the moment you set ANTHROPIC_API_KEY / OPENAI_API_KEY.
• Real memory, not a vector dump - LLM extraction distills conversations into discrete facts; reconciliation deduplicates, merges, and supersedes contradicted memories instead of deleting them (bi-temporal: valid_from / invalid_at / superseded_by).
• Explainable retrieval - hybrid vector + BM25 with reciprocal-rank fusion, boosted by recency and importance; every result carries its score signals.
• Provenance & audit - raw episodes are stored immutably; every memory links to its source episodes; every mutation is an event you can inspect (history).
• Forgetting built in - importance decays with a half-life; a decay sweep soft-forgets stale trivia (invalidated, never destroyed).
• Entities, disambiguated - mentions become first-class entities, and a name match is never enough to merge: unclear cases stay separate ("three Jonases") with a merge proposal you (or the system, once evidence is clear) confirm or reject later.
• Category filters - search(categories=["diet"]), memry search -c diet, ?categories= on REST, categories on the MCP search tool.
• Scales when you need it - optional usearch HNSW index (memry[ann]) for vector search at scale, and a PostgreSQL + pgvector backend (memry[postgres]) for multi-writer deployments. SQLite stays the zero-ops default.
• Multi-tenant ready - per-tenant API keys with transparent namespacing and strict isolation on the self-hosted server (MEMRY_TENANTS), plus an admin key.
• Research-grade - a built-in eval harness (recall@k, MRR, latency) with a synthetic dataset, plus a pluggable backend interface with an optional Mem0 adapter so you can benchmark against it under identical conditions.

Apache-2.0. Your agents. Your memories. Your infrastructure.

---

Quickstart

1. As an MCP server (any agent)

// Claude Desktop / Cursor / Windsurf config
{
  "mcpServers": {
    "memry": {
      "command": "memry",
      "args": ["mcp"],
      "env": { "ANTHROPIC_API_KEY": "sk-ant-..." }   // optional but recommended
    }
  }
}

# Claude Code
claude mcp add memry -- memry mcp

The server exposes: save_memories, search_memories, get_memory_context, list_memories, update_memory, delete_memory, memory_history, memory_stats.

2. As a Python library

from memry import MemoryStore

store = MemoryStore()

# write: extraction + reconciliation (or infer=False to store verbatim)
store.add(
    [{"role": "user", "content": "I'm Ada, a data engineer in Berlin. I prefer uv over pip."}],
    user_id="ada",
)

# read: hybrid search with explainable scores
for hit in store.search("what tooling does the user prefer?", user_id="ada"):
    print(hit.score, hit.memory.content, hit.signals)

# or a ready-to-inject, token-budgeted context block
ctx = store.reconstruct_context("help me set up a new project", user_id="ada", token_budget=1200)
print(ctx.text)

3. Self-hosted server (REST + dashboard + MCP)

memry serve --host 0.0.0.0 --port 8787
# dashboard:  http://localhost:8787/
# REST API:   http://localhost:8787/api/v1/...
# MCP (HTTP): http://localhost:8787/mcp

Or with Docker:

docker compose up -d      # see docker-compose.yml

Or on a fresh VPS (Ubuntu/Debian) with automatic HTTPS - one command, any provider (guide):

curl -fsSL https://raw.githubusercontent.com/cosmin-novac/memry/main/deploy/install.sh \
  | MEMRY_DOMAIN=memory.example.com bash

Set MEMRY_API_KEY to require Authorization: Bearer <key> on the API.

4. CLI

memry add "I moved to Amsterdam and joined ASML" -u ada
memry search "where does ada work" -u ada
memry context "plan a commute" -u ada
memry history <memory_id>          # full audit trail
memry sweep                        # decay: soft-forget stale memories
memry eval --dataset evals/datasets/synthetic_v1.jsonl

---

How it works

flowchart LR
    A[conversation] --> E[episodes<br/><i>immutable raw log</i>]
    A --> X[extraction<br/><i>LLM distills facts</i>]
    X --> R{reconcile}
    R -- new --> ADD[ADD memory]
    R -- overlaps --> UPD[UPDATE in place]
    R -- contradicts --> SUP[invalidate old<br/>supersede with new]
    R -- duplicate --> NONE[skip]
    ADD & UPD & SUP --> M[(memories<br/>FTS5 + vectors + events)]
    Q[agent query] --> H[hybrid retrieval<br/>RRF + recency + importance]
    M --> H --> C[token-budgeted<br/>context block]

1. Episodes first. Every message is stored verbatim before anything is derived from it. Memories are an index; episodes are the source of truth - you can re-run a better extraction pipeline over them later.
2. Extraction (Mem0-style phase 1): an LLM distills discrete, self-contained facts with type (semantic / episodic / procedural), importance, categories, and entities. Without an LLM, messages are stored verbatim so the system still works.
3. Reconciliation (phase 2, with Zep-style temporal semantics): each fact is compared to its most similar existing memories - duplicates are skipped, refinements rewrite in place, and contradictions invalidate the old memory and link it to its successor.
4. Retrieval: BM25 + cosine similarity fused with RRF, then boosted by recency (half-life) and importance. Keyword-only when no embedder is configured.
5. Forgetting: effective importance decays over time; memry sweep invalidates memories that decayed below threshold.

Configuration

Everything works with defaults. Override via env vars, ~/.memry/config.json, or Config(...):

Env var	Default	Notes
MEMRY_DB_PATH	~/.memry/memry.db	single SQLite file
MEMRY_BACKEND	local	local | mem0 (needs memry[mem0])
MEMRY_DEFAULT_USER	default	user scope when the agent doesn't pass one
MEMRY_LLM_PROVIDER	auto	anthropic | openai | ollama | none - auto-detected from ANTHROPIC_API_KEY / OPENAI_API_KEY
MEMRY_LLM_MODEL	per provider	claude-opus-4-8 / gpt-5-mini / llama3.1 (use claude-haiku-4-5 for a cheaper extraction model)
MEMRY_EMBEDDING_PROVIDER	auto	openai | ollama | voyage | hash | none
MEMRY_API_KEY	-	bearer token for the REST/MCP HTTP server

Anthropic extraction requires the optional SDK: pip install "memry[anthropic]".

Evaluation

memry eval --dataset evals/datasets/synthetic_v1.jsonl -k 5

The harness ingests each case through the full write path, then scores retrieval (recall@k, MRR, latency p50/p95) - deterministic and offline, so it runs in CI. Format LoCoMo/LongMemEval into the same JSONL schema to compare providers, configs, and backends (including the Mem0 adapter) under identical conditions. See docs/research/competitive-analysis.md for the landscape survey behind the design.

Project layout

src/memry/
  models.py            # Episode / Memory / MemoryEvent (bi-temporal, provenance)
  config.py            # env + file config, provider auto-detection
  store.py             # MemoryStore - the public API
  retrieval.py         # hybrid search: RRF + recency + importance
  backends/            # storage interface, local SQLite engine, Mem0 adapter
  intelligence/        # extraction, reconciliation, decay, context building
  providers/           # LLMs (Anthropic/OpenAI/Ollama) & embeddings (+hash fallback)
  mcp_server.py        # MCP tools (stdio + streamable HTTP)
  rest.py              # REST API + dashboard + /mcp mount
  evals/               # retrieval eval harness

Development

pip install -e ".[dev]"
pytest

License

Apache-2.0