Memry

Memry is the long-term memory system for AI agents. It is open source and you host it yourself - memry.tech

pip install memry
export OPENAI_API_KEY=sk-... MEMRY_DECISION_PROVIDER=jev MEMRY_DECISION_API_KEY=...
memry mcp        # your agent now has long-term memory

Memry gives any MCP-capable agent - Claude Code, Claude Desktop, Cursor, Windsurf, Codex - durable long-term memory. It distills conversations into discrete facts, reconciles each new fact against what it already knows, and serves the result back as token-budgeted context. All knowledge state is a single SQLite file on your machine: no external vector database or queue service. To run it, you need two models: a text model (OpenAI, Anthropic or Ollama) and a decision model, TypeSafe Jev.

A scratchpad or a MEMORY.md file is text your agent rereads at the start of every session. You keep it short by hand, and a line that is no longer true is still in the file until you delete it. Memry is a database that holds every fact your agents have saved and returns the few that fit the question. Every MCP client you connect reads and writes the same memory.

Why Memry

It runs anywhere, with two models. Knowledge storage is one SQLite file with no services around it. A text model extracts the facts from what you save. A decision model, Jev, returns a calibrated probability that two names belong to one thing, and Memry merges the duplicates on its own. The server doesn't start without both models. If you set MEMRY_DECISION_PROVIDER=llm, Memry sends the decision questions to the text model, merges entities only by fixed rules, and you confirm the other merges yourself.

It remembers the way you would want it to. New facts are reconciled against existing ones: duplicates are skipped, refinements update in place, and contradictions supersede the old memory instead of deleting it. Memories are bi-temporal (valid_from / invalid_at / superseded_by), so "moved to Amsterdam" does not erase "lived in Berlin" - it dates it. Nothing is forgotten for being old: a memory leaves use only when you delete it or a later one replaces it, and either can be undone.

Nothing is a black box. Raw episodes are stored immutably before anything is derived from them, every memory links back to its source episodes, every mutation is an inspectable event, and every search hit carries its score signals (BM25, vector, recency, importance). When a memory looks wrong, you can trace where it came from, or re-run a better extraction pipeline over the original episodes.

Entities are disambiguated by evidence. Mentions become first-class entities. An exact first-and-last name with overlapping context merges automatically; a shared short name or full name without supporting context stays separate with a reviewable merge proposal. Prior merges are followed, so an older proposal cannot leave a broken merge target.

Topics organize themselves. Extraction tags each memory at the level a later conversation would ask about (liver health, weekly gym, 2026 taxes), reusing the vocabulary already in your store rather than coining a synonym every session. Mechanical variants such as food/foods merge without review, and two tags that name one subject are compared and merged as two names of one thing are. Search and list by topic or by date window.

It stays simple as one shared service. SQLite is the sole production store. One Memry server can serve many agents, devices, and tenant namespaces without an external database. For larger stores, the optional memry[ann] extra adds a rebuildable usearch HNSW candidate index. Multiple server replicas writing one store are not currently supported.

Multi-user, with real OAuth. Beyond static config tenants, one server can host runtime-managed accounts, created with memry account add. The first account is the bootstrap administrator and keeps only the existing default memory space; every later account gets one private memory space. Set MEMRY_PUBLIC_URL and Memry becomes an OAuth 2.1 authorization server for those accounts (dynamic client registration, PKCE, refresh, revocation, discovery at the domain root), so any OAuth-capable MCP client - Claude, Cursor, VS Code - can sign in and get a token scoped to that account. No IdP required; Memry verifies against its own accounts. Off by default: the single-user path stays keyless. See docs/self-hosting.md.

You can measure it. A built-in eval harness scores retrieval (recall@k, MRR, latency percentiles) deterministically and offline. An optional Mem0 adapter lets comparison or import tooling read and exercise Mem0 under the same interface; it cannot be selected as Memry's runtime store.

Quickstart

As an MCP server (any agent)

Memry speaks MCP two ways: stdio for agents on the same machine (zero config, no port, no auth) and streamable HTTP for a shared server that several agents and devices talk to.

Local, stdio - the fastest start:

# Claude Code
claude mcp add memry -e OPENAI_API_KEY=sk-... -e MEMRY_DECISION_PROVIDER=jev \
  -e MEMRY_DECISION_API_KEY=... -- memry mcp

// Claude Desktop / Cursor / Windsurf config
{
  "mcpServers": {
    "memry": {
      "command": "memry",
      "args": ["mcp"],
      "env": {
        "OPENAI_API_KEY": "sk-...",           // text model (or ANTHROPIC_API_KEY)
        "MEMRY_DECISION_PROVIDER": "jev",     // decision model
        "MEMRY_DECISION_API_KEY": "..."       // TypeSafe key
      }
    }
  }
}

The Anthropic SDK is an optional extra: pip install "memry[anthropic]". Without it an ANTHROPIC_API_KEY is ignored with a warning, and with no other text model the server does not start. With OPENAI_API_KEY you don't need the extra.

Remote, streamable HTTP - point any MCP client at a self-hosted server (see below) and share one memory across every machine:

# Claude Code
claude mcp add --transport http memry https://memory.example.com/mcp \
  --header "Authorization: Bearer <MEMRY_API_KEY>"

// Cursor / Windsurf / anything that takes a config file
{
  "mcpServers": {
    "memry": {
      "type": "http",
      "url": "https://memory.example.com/mcp",
      "headers": { "Authorization": "Bearer <MEMRY_API_KEY>" }
    }
  }
}

claude.ai (web, desktop, mobile) - add Memry as a custom connector under Settings → Connectors → Add custom connector. The dialog has no header field, so embed the key in the URL instead:

https://memory.example.com/mcp/<MEMRY_API_KEY>

Full walkthrough with screenshots of the flow, security notes, and troubleshooting: docs/connect-claude-ai.md.

ChatGPT - add Memry as a connector and sign in with OAuth (no key to paste). Set MEMRY_PUBLIC_URL first, then give ChatGPT the MCP URL https://memory.example.com/mcp; the bare origin works too. Walkthrough and troubleshooting: docs/connect-chatgpt.md.

The server exposes save_memories, search_memories, get_memory_context, list_memories, list_categories, update_memory, delete_memory, memory_history, and memory_stats. Recall takes two filters, applied before ranking: when (a day, month, year or range: "Was habe ich am 01. April 2025 gemacht?" is when="2025-04-01") and about (names of people, projects, things or tags); a phrase in double quotes in the query must appear exactly. Agents are instructed to recall context at the start of a task and to send save_memories what was said in words close to the original, one statement per line, with the speaker named when it is someone other than the user ("Ada: I got the job"). Memry extracts the facts and keeps the saved text as the source turns it shows with each memory in later searches. If an agent sends a summary, those searches show the summary, and the details it dropped are lost. Related statements go in one call. If they arrive in separate calls, the client can repeat a semantic context label and run_id; up to three optional tags are treated as classification hints. For something said on another day, such as an imported chat, the client passes said_at (YYYY-MM-DD): the memories are dated that day, and "yesterday" or "next month" in the text counts from it. When a fact changes or the user corrects it, the agent saves the new statement, and Memry keeps the old value as dated history or retires it when it was wrong. With the default infer=true, the exact text is acknowledged immediately and remains searchable. The managed worker waits for two minutes of quiet, then distills each related group when an LLM is configured.

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)

As a self-hosted server (REST + dashboard + MCP)

export OPENAI_API_KEY=sk-... MEMRY_DECISION_PROVIDER=jev MEMRY_DECISION_API_KEY=...
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

The dashboard shows your memories with inline editing, filtered search, lossless JSON backup/restore, a unified Upkeep area, and a galaxy map aggregated over every active memory independently of the paginated detail list. The map groups by entity: it shows hubs, with the parts of a project or product as moons on it, and tags once their type is turned on; concept, other and tag are hidden by default and the type menu controls what is shown. Upkeep > Entities lists people, things and tags in one list filtered by type, and the memory list's About filter picks any of them. Heavily-used groups gravitate to the gold core, the working set orbits in the teal belt, and one-off groups drift at the violet rim. Orbit-marker shapes distinguish semantic, procedural, episodic, and working memories. Idle link and orbit rendering is bounded for large stores (above 400 groups, orbit markers stay on the core and on whatever you hover or select); selecting a planet reveals its complete visible neighborhood and filters the detail list through the server. Selecting an entity also surfaces its summary, aliases, and rename control, with explicit controls to merge a duplicate or remove a mistaken entity without deleting memories. When a mistaken entity occurs in multiple memories, its name is retained as a tag. Opening Upkeep temporarily unloads the map and restores it on close to avoid holding both views in memory.

With Docker: docker compose up -d --build (see docker-compose.yml).

Or on a fresh Ubuntu/Debian VPS, one command installs Docker, Memry, and Caddy with automatic HTTPS (full 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. Without a key (or accounts) the server only binds loopback; MEMRY_ALLOW_OPEN=1 overrides that when a reverse proxy protects the port. More in docs/self-hosting.md. To plug a hosted Memry into claude.ai as a custom connector, see docs/connect-claude-ai.md.

Multi-user accounts

One server can host multiple accounts. The first account is the bootstrap administrator and uses only the existing default memory space. Every later account uses only its own name::default space. The role does not expose other accounts' memories. Manage accounts on the server with the CLI:

memry account add alice --password s3cret    # creates it, prints an API key (shown once)
memry account list
memry account issue-key alice --label laptop # another key for the same account
memry account disable alice                   # its keys + sessions stop working immediately

An account connects with its API key (Authorization: Bearer <key>, or https://<host>/mcp/<key>), or - once you set MEMRY_PUBLIC_URL=https://memory.example.com - by signing in through OAuth from any client that supports it (Claude, ChatGPT, Cursor, VS Code). On the dashboard, accounts sign in at /login with their name and password. Full walkthrough: docs/self-hosting.md.

From the 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 eval --dataset evals/datasets/synthetic_v1.jsonl

How it works

flowchart LR
    A[conversation] --> E["episode + active pending memory"]
    E --> ACK[durable acknowledgement]
    E --> X[background extraction]
    X --> R{reconcile}
    R -->|new| ADD[add]
    R -->|refines| UPD[update in place]
    R -->|contradicts| SUP[supersede old]
    R -->|duplicate| SKIP[skip]
    ADD --> M[(memories)]
    UPD --> M
    SUP --> M
    Q[agent query] --> H[hybrid retrieval]
    E --> H
    M --> H
    H --> C[token-budgeted context]

1. Durable acknowledgement. The MCP fast-save path commits the exact input as an immutable episode and active, searchable pending memory before replying. That SQLite row is also the recovery marker, so no external queue is required.
2. Managed enrichment. One in-process worker drains small database batches. Each payload is extracted separately so user scopes, provenance, and retries cannot mix. Provider failure leaves the raw memory active and schedules a bounded-backoff retry; restart recovery reads the same pending rows. Without an LLM key, they stay verbatim.
3. Reconciliation. Each extracted fact is compared to similar active memories: duplicates are skipped, refinements rewrite in place, contradictions invalidate the old memory and link it to its successor. The pending raw memory is superseded only after enrichment succeeds.
4. Retrieval. BM25 and cosine similarity are fused with reciprocal-rank fusion, then boosted by recency and importance. A question naming a known entity also takes the memories of the entities linked to it (the linked search). Pending raw memories are searchable immediately.
5. No forgetting by age. Recency lifts newer memories in ranking, but age never takes a memory out of use. Tag filters (memry search -c diet) narrow any query.

Configuration

Before you start a server, set a text model and a decision model; the server doesn't start without them (docs/self-hosting.md). Every other setting has a default. Override via env vars, ~/.memry/config.json, or Config(...):

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

Memry sends its decision questions to TypeSafe Jev

Memry's decision questions each have a fixed set of answers: are these two entities the same one, how long does this fact stay worth remembering, do these memories contain the same fact, which result is the best answer to the question. With MEMRY_DECISION_PROVIDER=jev, Memry sends them to TypeSafe Jev, a hosted System One model that returns a calibrated probability for each possible answer.

export MEMRY_DECISION_PROVIDER=jev
export MEMRY_DECISION_API_KEY=...   # TypeSafe API key

With Jev, Memry merges duplicate entities on its own above 0.70 confidence, a threshold measured on the labelled identity set in evals/. Jev also judges the first 20 results of a search, and for a question whose answer is a set (a list, a total, a comparison) one more batch of memories. Scoring how long each memory stays relevant is off unless MEMRY_DURABILITY=1. You still need a text model for extraction. If you set MEMRY_DECISION_PROVIDER=llm, Memry sends these questions to the text model, merges entities only by fixed rules, and you confirm the other merges yourself. The measurements and the remaining settings are in docs/self-hosting.md.

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). It is deterministic and offline, so it runs in CI. Datasets in its JSONL schema compare providers, configs and the optional Mem0 comparison adapter under identical conditions. evals/external_benchmarks.py runs LoCoMo and LongMemEval as published (evidence recall at 5, 10 and 20 and MRR, without a model).

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
  enrichment.py        # managed pending-memory worker and restart recovery
  retrieval.py         # hybrid search: RRF + recency + importance
  backends/            # storage contract + the production SQLite engine
  intelligence/        # extraction, reconciliation, entities, 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; supports explicit comparison adapters

Development

pip install -e ".[dev]"
pytest

License

Apache-2.0