RE-call MCP Memory Server

[RE-call: memory that knows when not to guess]

Memory that abstains instead of guessing.
RE-call is agent memory in a local SQLite file or your own PostgreSQL with pgvector. Each result carries a verdict, confidence, and provenance, while unsupported questions are refused instead of answered by the nearest match.

[CI] [PyPI version] [License: AGPL-3.0] [Python 3.11+] [PostgreSQL + pgvector]

What it is  ·  For agents  ·  How it works  ·  Start with Lite  ·  Install and integrate

What is RE-call

RE-call is agent memory in a local SQLite file or your own PostgreSQL database. It indexes source documents and keeps validity, lineage, confidence, and provenance attached to every result.

Vector search returns nearby text. RE-call also checks whether that text is current, supported, and trustworthy enough for the query. A superseded claim is marked superseded; a result that does not clear the calibrated trust gate becomes ABSTAIN with a reason. Declared supersession makes the current memory win over stale but similar memory.

The default path uses local embeddings plus hybrid dense and keyword retrieval. PostgreSQL uses its full text search; Lite uses SQLite FTS5. It needs no memory-layer LLM call. Additional retrieval, reasoning, and structured-fact modules are optional.

For the agent reading this, RE-call for agents explains the experience loop and how to use persistent memory as a trustworthy past rather than an unverified scratchpad.

How it works

On PostgreSQL, RE-call builds an immutable, calibrated generation and serves queries against that pinned generation. Lite binds calibration to its single-file corpus. Every search, including the reasoning and fact paths, goes through trusted search and the trust gate.

[RE-call architecture: entry points, the generation build, trusted retrieval against a pinned generation, and the trust gate]

The diagram is interactive on the project site: click any box to see the source lines it describes. A step by step view of one recall_search call follows a single search through its refusal, abstention and answer paths. Both are generated from the sources in docs/diagrams.

flowchart LR
subgraph group_entry["Entry points"]  node_cli["recall CLI<br/>commands"]  node_mcp["MCP server<br/>tool handlers"]  node_sdk["Agent SDK<br/>in process"]end
subgraph group_build["Generation build"]  node_manifest["Source manifest<br/>immutable input"]  node_generations["Generation manager<br/>build, validate, promote"]  node_parser["Parse and chunk<br/>documents to chunks"]  node_embedder["Embedder<br/>local or hosted API"]  node_graph_build["Semantic graph<br/>entities, relations"]  node_calibration["Calibration<br/>certified threshold"]end
subgraph group_serve["Trusted retrieval"]  node_trusted_search["Trusted search<br/>query orchestrator"]  node_gen_store["Generation store<br/>pinned generation"]  node_retriever["Hybrid retriever<br/>dense, full text, RRF"]  node_reranker["Reranker<br/>opt in"]  node_trust_gate{{"Trust gate<br/>verdict or ABSTAIN"}}end
subgraph group_reason["Reasoning"]  node_reasoner["Reasoning query<br/>evidence bundle, plan"]  node_graph_expansion["Graph expansion<br/>one hop, before trust"]  node_answer["Answer provider<br/>opt in LLM"]end
subgraph group_facts["Provenance and facts"]  node_provenance["Provenance controller<br/>fact review"]  node_ledger["Fact ledger<br/>append only"]end
node_caller(("Agent or user<br/>caller"))node_postgres[("PostgreSQL + pgvector<br/>generations, chunks, graph")]
node_caller -->|"runs"| node_clinode_caller -->|"tool calls"| node_mcpnode_caller -->|"imports"| node_sdknode_sdk -->|"reuses handlers"| node_mcpnode_cli -->|"generation build"| node_generationsnode_mcp -->|"ingest uploads"| node_generationsnode_generations -->|"reads sources"| node_manifestnode_generations -->|"parses, chunks"| node_parsernode_generations -->|"embeds chunks"| node_embeddernode_generations -->|"builds graph"| node_graph_buildnode_generations -->|"requires certified"| node_calibrationnode_generations -->|"writes generation"| node_postgresnode_graph_build -->|"writes graph"| node_postgresnode_cli -->|"calibrate, publish"| node_calibrationnode_calibration -->|"stores threshold"| node_postgresnode_cli -->|"search"| node_trusted_searchnode_mcp -->|"search, evidence"| node_trusted_searchnode_trusted_search -->|"pins generation"| node_gen_storenode_gen_store -->|"reads active"| node_postgresnode_trusted_search -->|"retrieves"| node_retrievernode_retriever -->|"embeds query"| node_embeddernode_retriever -->|"dense, full text"| node_gen_storenode_retriever -.->|"reorders"| node_rerankernode_trusted_search -->|"evaluates"| node_trust_gatenode_mcp -->|"reasoning query"| node_reasonernode_reasoner -->|"retrieves"| node_trusted_searchnode_trusted_search -.->|"expands pre trust"| node_graph_expansionnode_graph_expansion -.->|"same generation graph"| node_postgresnode_reasoner -.->|"cited answer"| node_answernode_mcp -->|"apply fact"| node_provenancenode_provenance -->|"fresh search"| node_trusted_searchnode_provenance -->|"appends"| node_ledgernode_ledger -->|"stores"| node_postgres
click node_cli "https://github.com/GiulioDER/RE-call/blob/master/recall/cli.py"click node_mcp "https://github.com/GiulioDER/RE-call/tree/master/recall_mcp"click node_sdk "https://github.com/GiulioDER/RE-call/blob/master/recall_agent/memory.py"click node_manifest "https://github.com/GiulioDER/RE-call/blob/master/recall/manifest.py"click node_generations "https://github.com/GiulioDER/RE-call/blob/master/recall/generations.py"click node_parser "https://github.com/GiulioDER/RE-call/blob/master/recall/document.py"click node_embedder "https://github.com/GiulioDER/RE-call/blob/master/recall/embeddings.py"click node_graph_build "https://github.com/GiulioDER/RE-call/blob/master/recall/semantic_graph/_build.py"click node_calibration "https://github.com/GiulioDER/RE-call/blob/master/recall/calibration_v2.py"click node_trusted_search "https://github.com/GiulioDER/RE-call/blob/master/recall/trust.py"click node_gen_store "https://github.com/GiulioDER/RE-call/blob/master/recall/generation_store.py"click node_retriever "https://github.com/GiulioDER/RE-call/blob/master/recall/retriever.py"click node_reranker "https://github.com/GiulioDER/RE-call/blob/master/recall/rerank.py"click node_trust_gate "https://github.com/GiulioDER/RE-call/blob/master/recall/trust.py"click node_reasoner "https://github.com/GiulioDER/RE-call/blob/master/recall/reasoning.py"click node_graph_expansion "https://github.com/GiulioDER/RE-call/blob/master/recall_mcp/graph_expansion.py"click node_answer "https://github.com/GiulioDER/RE-call/blob/master/recall/answer_provider.py"click node_provenance "https://github.com/GiulioDER/RE-call/blob/master/recall/provenance_controller.py"click node_ledger "https://github.com/GiulioDER/RE-call/blob/master/recall/fact_ledger.py"click node_postgres "https://github.com/GiulioDER/RE-call/tree/master/recall/migrations"
classDef toneNeutral fill:#f8fafc,stroke:#334155,stroke-width:1.5px,color:#0f172aclassDef toneBlue fill:#dbeafe,stroke:#2563eb,stroke-width:1.5px,color:#172554classDef toneAmber fill:#fef3c7,stroke:#d97706,stroke-width:1.5px,color:#78350fclassDef toneMint fill:#dcfce7,stroke:#16a34a,stroke-width:1.5px,color:#14532dclassDef toneRose fill:#ffe4e6,stroke:#e11d48,stroke-width:1.5px,color:#881337classDef toneIndigo fill:#e0e7ff,stroke:#4f46e5,stroke-width:1.5px,color:#312e81classDef toneTeal fill:#ccfbf1,stroke:#0f766e,stroke-width:1.5px,color:#134e4aclass node_cli,node_mcp,node_sdk,node_caller toneBlueclass node_manifest,node_generations,node_parser,node_embedder,node_graph_build,node_calibration,node_postgres toneAmberclass node_trusted_search,node_gen_store,node_retriever,node_reranker,node_trust_gate toneMintclass node_reasoner,node_graph_expansion,node_answer toneRoseclass node_provenance,node_ledger toneIndigo

Ordinary recall search follows the direct path. Explicit reasoning accepts graph_expansion: auto is the default and resolves to bounded one hop expansion for every nonempty query; off keeps direct retrieval only; one-hop forces the graph path. The CLI uses --graph-expansion auto|off|one-hop; the MCP tool uses graph_expansion="auto"|"off"|"one_hop". Graph neighbors are generation bound, direct candidates remain first, and expanded candidates must clear the same trust boundary before they can support a cited answer.

The opt in choices attach to different points in the system:

For details, see the architecture writeup, provenance controller, and API reference.

Quickstart

For one project's memory on your machine, start with RE-call Lite. It stores memory in .recall/memory.db, refreshes changed memos automatically, and needs no Docker or database server.

Install the source build first: Lite is in the repository; the published 0.14.0 PyPI wheel predates it. Follow the Lite installation guide to create an environment and install the source, then run this from your own project with an existing folder of notes:

recall setup --lite --memory ./notes

Setup indexes the notes, attempts automatic calibration, and registers Claude Code for this project. Inspect the calibration and registration results, use the printed doctor command, and restart the client. Strict trust stays enabled; a corpus that cannot certify will refuse trusted answers. For other MCP clients, use manual Lite wiring.

Try the PostgreSQL demo

Prerequisites: Python 3.11 or newer, Docker, and a Docker installation able to run PostgreSQL with pgvector.

pip install "recall-rag[fastembed]"recall quickstart

The demo starts a throwaway database, indexes a small corpus included in the package, and runs three searches. It includes a normal answer, a stale claim that is returned as superseded, and a question that is refused. The demo uses development trust and changes no personal files.

Remove the demo database when finished:

recall quickstart --remove

Already have PostgreSQL with pgvector? Use recall quickstart --existing-dsn <dsn> instead. The demo is intentionally separate from a real install and is not calibrated for your data.

Install and integrate

Local, with no database server and no Docker

Use the Lite source installation, then run from the project whose notes you want to index:

recall setup --lite --memory ./notes

Setup creates .recall/memory.db, attempts automatic calibration, and registers Claude Code for this project. Use the printed doctor and read-only dashboard commands, and add .recall/ to .gitignore. For CLI or manual client wiring, add --no-register.

The MCP server refreshes changed memos automatically. To check or refresh explicitly:

recall lite refresh --store .recall/memory.db --memory ./notes --checkrecall lite refresh --store .recall/memory.db --memory ./notes

Use the same memo root and embedder as setup. The CLI does not read the MCP client's environment; apply setup's printed settings before CLI searches or doctor. Strict trust remains enabled, and successful indexing alone does not guarantee certification.

The Lite guide covers other clients, source updates, backup and restore, refresh budgets, and queued writes. Keep the file on a local disk. Lite uses exact cosine search and SQLite keyword retrieval; results can differ from PostgreSQL. Generations, multi-tenant serving, and authentication require PostgreSQL. See the API for trust and retrieval limits.

With PostgreSQL

For your own corpus, provide PostgreSQL with pgvector and run the guided setup wizard after installing recall-rag[fastembed]:

recall setup

It applies the schema, asks for the embedder and retrieval options, indexes the corpus, offers calibration, and registers the selected agent integration. When the wizard asks whether to calibrate, use a labeled query file that refers to the corpus you are installing. Calibration fitted to the bundled demo is only an example, not a certification for your data. The schema uses an ordered SQL migration path and pre-tenancy tables are migrated in place.

For Docker, an existing database, headless provisioning, manual calibration, and troubleshooting, see docs/INSTALLATION.md and docs/WIZARD.md.

Choose an integration

Claude Code

Inside Claude Code, install the plugin after installing the Python package:

/plugin marketplace add GiulioDER/RE-call/plugin install recall@re-call

The plugin supplies the MCP server, memory search skill, and lifecycle hooks. Lite setup already registers a project server; the plugin is optional. If you add it, configure its Database DSN with the SQLite DSN printed by setup and avoid enabling duplicate servers for the same memory. The plugin's server does not refresh memos by itself; run recall lite refresh yourself or keep the server setup registers. For PostgreSQL, run recall setup against the project and database Claude should use. Details are in plugin/README.md.

Codex

For Lite, use manual MCP wiring: recall setup --lite only registers Claude Code. For PostgreSQL, run recall setup from the project. When Codex is detected, setup installs the Codex MCP server, plugin bundle, memory and session skills, and hooks into the user configuration. Restart Codex afterward. The Codex and Claude Code integrations share the same memo format and trust layer. See docs/CODEX_RECALL_INTEGRATION.md.

Claude Agent SDK

The SDK integration runs the same tools in process and does not start an MCP server:

from recall_agent import RecallAgentMemory
with RecallAgentMemory.from_env() as memory:    options = memory.options()

See docs/USING_WITH_AGENT_SDK.md for the complete example and write-tool boundaries.

LangChain and LlamaIndex

Both adapters use the same trusted retrieval path. If trust abstains, they return no document by default, and returned documents retain verdict, confidence, cosine, and supersession metadata. See docs/API.md for the supported classes and methods.

If an install is not working

recall doctor

For Lite, first apply the environment printed by setup so doctor inspects the SQLite store. For PostgreSQL, the doctor checks the interpreter, console scripts, embedder, Docker, database, pgvector, schema, configured table and tenant, calibration, and agent registration. It changes nothing and prints the repair command for each problem.

Product surface

RE-call is not a hosted memory service, dashboard, or automatic truth extractor. It does not rewrite corpus metadata from an agent's inference. Reasoning is opt in, citation constrained, and review aware. See docs/PRODUCTION.md for deployment boundaries.

Read next

Published numbers are tied to committed artifacts, and the claim gate checks them in CI. Benchmark interpretation and limits belong in docs/EVIDENCE.md, not in this overview.

Citation

If you describe RE-call in a paper, post, talk, or README of your own, cite the project and credit Giulio D'Erme. Use CITATION.cff as the canonical citation source.

License

RE-call is open source under the GNU Affero General Public License v3.0 (AGPL-3.0). You may use, study, modify, and share it, commercially as well, on the AGPL's conditions: if you distribute RE-call or a work based on it, or let users interact with a modified version over a network, you must make the corresponding source available under the same license. When those conditions do not fit, for example a closed source product or a proprietary hosted service, a commercial license is available from the copyright holder. See COMMERCIAL_LICENSE.md for the boundary between the two and for the terms of earlier releases, and preserve NOTICE when redistributing the software. Contributions are accepted under the Contributor License Agreement.