recall-rag 0.9.1

Trustworthy memory for AI agents.
RE-call gives retrieval results confidence, provenance, validity, tenant isolation, and an explicit abstention path when the memory does not support an answer.

Why RE-call  ·  Five-minute proof  ·  Quickstart  ·  How it works  ·  Product surface  ·  Documentation  ·  Evidence

Why RE-call

Most memory systems optimize for the nearest match. Agent memory needs a stricter contract: the retriever must say whether a memory is current, where it came from, how confident it is, and when the corpus does not contain an answer.

RE-call is built around that contract.

It is for teams putting agent memory behind real applications: support copilots, internal research agents, compliance assistants, and long-running workflow agents where a stale or unsupported memory is worse than no memory. The buyer story is simple: keep the memory layer local by default, attach policy to every hit, calibrate the refusal threshold on your corpus, and let the application decide what to do with a result that is not trustworthy enough to answer from.

Measured strengths:

The README is the product overview. For evidence behind these claims, start with docs/EVIDENCE.md, then use results/FINDINGS.md for the full interpretation and limits.

Five-minute proof

Run the bundled demo to see the product behavior before reading the evidence docs:

docker compose up -d --wait
pip install "recall-rag[fastembed]"
python -m recall.cli --table recall_quickstart \
  --migration-dsn postgresql://recall:recall@localhost:5432/recall \
  schema --dim 384 apply
RECALL_TRUST_MODE=development python -m recall.cli --table recall_quickstart demo

Expected shape:

[DEGRADED:INDEX_NOT_READY] query='how many requests per second can a client make?'
  ok          conf=1.00  cos=0.784  rate_limits_v2.md
  superseded  conf=1.00  cos=0.806  rate_limits_v1.md -> use rate_limits_v2.md

[ABSTAIN GAP DEGRADED:INDEX_NOT_READY] query='how do we handle penguins on mars?'
  reason: no hit above the calibrated confidence threshold

The stale memory is more similar to the query, but it is declared superseded and loses to the current memory. The unrelated query returns an abstention. The degraded marker is intentional here: this is a sample-corpus demonstration, not a certified production calibration.

Runnable examples: examples/README.md.

Quickstart

After the demo, run the guided setup wizard for your own corpus. The wizard records the selected embedder, retrieval options, and an optional calibration that is fitted to your labeled queries and your corpus.

docker compose up -d --wait
pip install "recall-rag[fastembed]"
python -m recall.cli --table recall_quickstart \
  --migration-dsn postgresql://recall:recall@localhost:5432/recall \
  schema --dim 384 apply
python -m recall.cli setup

PowerShell:

docker compose up -d --wait
pip install "recall-rag[fastembed]"
python -m recall.cli --table recall_quickstart `
  --migration-dsn postgresql://recall:recall@localhost:5432/recall `
  schema --dim 384 apply
python -m recall.cli setup

When the wizard asks whether to calibrate, provide a labeled query JSON and the corpus directory. Use recall/eval/queries.json as the input shape. Calibration is per embedder and per corpus, so a new model or substantially changed corpus should be calibrated again.

The distribution is recall-rag; the import is recall. The name recall on PyPI belongs to an unrelated package, so do not install both into the same environment.

Working from a clone:

pip install -e ".[fastembed]"

How it works

flowchart TB
    M["Memo: markdown plus frontmatter"] --> CH["Chunk"]
    CH --> EW["Embed locally"]
    EW -. "optional" .-> SP["SPLADE encode"]
    EW --> DB
    SP -. "optional" .-> DB

    Q["Query"] --> EQ["Query encoder"]
    EQ --> DB[("PostgreSQL plus pgvector")]

    DB --> DN["Dense vector search"]
    DB --> SL["Postgres full-text search"]
    DB -. "optional" .-> LS["Learned sparse search"]

    DN --> F["Reciprocal Rank Fusion"]
    SL --> F
    LS -. "optional" .-> F

    F -. "optional" .-> RR["Cross-encoder rerank"]
    RR --> GP
    F --> GP{"Gap check: calibrated threshold"}
    GP --> TR{"Trust layer: supersession, validity, confidence"}
    CAL["Calibration: fitted per embedder and corpus"] --> TR
    TR -. "optional" .-> EJ{"Entailment judge"}
    EJ --> OUT
    TR --> OUT["Verdict, confidence, provenance, or ABSTAIN"]

Product surface

Deliberately out of scope: an end-user dashboard, graph reasoning, entity synthesis, high availability orchestration, and automatic truth inference from prose.

The ordered SQL migration path is versioned now, pre-tenancy tables are migrated in place, and runtime CREATE TABLE IF NOT EXISTS remains bootstrap only.

Use it

For an ad hoc local markdown folder, create a table for that index, index the corpus, and search it. If you did not calibrate during setup, use development mode only for local evaluation and demos. Replace ./notes with your memo folder.

python -m recall.cli --table recall_notes \
  --migration-dsn postgresql://recall:recall@localhost:5432/recall \
  schema --dim 384 apply
RECALL_TRUST_MODE=development python -m recall.cli --table recall_notes index ./notes
RECALL_TRUST_MODE=development python -m recall.cli --table recall_notes search "what did we decide about caching?"
python -m recall.cli lint ./notes
python -m recall.cli check ./notes/new-memo.md --strict

PowerShell uses the same commands, but set development mode first when you are running an uncalibrated local evaluation:

$env:RECALL_TRUST_MODE = "development"

For production generation mode, build, validate, calibrate, and promote an immutable generation. Then query the tenant's active generation:

from recall.embeddings import FastEmbedEmbedder
from recall.generation_store import GenerationStore
from recall.trust import trusted_search

emb = FastEmbedEmbedder()
with GenerationStore(DSN, dim=emb.dim, tenant="acme", pool_size=8) as store:
    store.check_schema()
    result = trusted_search(store, emb, "what is the rate limit?")
    if result.abstained:
        ...  # say you do not know
    for hit in result.hits:
        hit.verdict
        hit.confidence
        hit.validity.superseded_by

Set RECALL_SERVING_DSN for application traffic and RECALL_MIGRATION_DSN only in the migration job. RECALL_DSN remains a deprecated development fallback for the serving DSN. See docs/MIGRATIONS.md. Configuration modes are summarized in docs/OPERATING_MODES.md.

Operational safety notes:

MCP

The MCP server uses the default chunks table. Apply that schema for the embedder the server will run, then point the client at recall_mcp.server.

python -m recall.cli --migration-dsn postgresql://recall:recall@localhost:5432/recall \
  schema --dim 384 apply

If an existing chunks table was created with another vector dimension, use a fresh database or an embedder with the matching dimension. The MCP stdio server does not take a --table flag.

{
  "mcpServers": {
    "recall": {
      "command": "python",
      "args": ["-m", "recall_mcp.server"],
      "env": {
        "RECALL_SERVING_DSN": "postgresql://...",
        "RECALL_TENANT": "acme",
        "RECALL_TRUST_MODE": "development"
      }
    }
  }
}

Omit RECALL_TRUST_MODE in production after you have built, calibrated, and promoted a generation. Local uncalibrated MCP work needs the explicit development setting for the same reason the CLI demo does.

Tools: recall_search, recall_evidence, recall_index, recall_forget, and recall_stats.

Full guide: docs/USING_WITH_CLAUDE.md. Authentication and tenancy: docs/AUTH.md.

LangChain and LlamaIndex

pip install "recall-rag[langchain]"
pip install "recall-rag[llamaindex]"

from recall.integrations.langchain import RecallRetriever

retriever = RecallRetriever.from_store(store, emb, k=5)
docs = retriever.invoke("what is the rate limit?")

When the trust layer abstains, the adapters return no document by default. Returned documents carry trust metadata, including verdict, confidence, cosine, and supersession details.

Documentation

Start with docs/README.md.

Core documents:

Release notes and upgrade warnings live in CHANGELOG.md.

Evidence

Start with benchmarks/README.md. The results directory has its own map at results/README.md.

The short version:

Important benchmark documents:

When not to use RE-call

Use something else if you need managed hosting, per-chunk ACLs, graph reasoning, automatic truth extraction from prose, or a memory system that rewrites facts for you. RE-call is a retrieval library over your PostgreSQL database, not a hosted memory platform.

What this does not do

RE-call is a retrieval library, not a general reasoning system. It does not infer every missing supersession edge, prove that an on-topic memory answers a near-miss question, or replace database operations with a managed service. It returns the trust signals the caller needs, and it refuses to pretend that a nearest match is always usable evidence.

Reproduce

make eval
python -m recall.eval.scale --embedder hashing --filler 50000

Cloud rows require the relevant API keys. Local rows run key-free.

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

Apache 2.0 license. See LICENSE, and keep NOTICE with redistributed derivative works.