recall-rag

Memory that abstains instead of guessing.
RE-call is agent memory on your own PostgreSQL with pgvector: every hit carries a verdict, confidence and provenance, a retracted claim comes back marked superseded, and a question the corpus cannot answer is refused rather than answered from the nearest neighbour.

ATM-Bench Recall@10 92.8924 against 79.09 for the best published row (limits)  ·  second of ten on MTRAG correct refusals (limits)  ·  zero memory-layer LLM calls to build memory, where Mem0 pays one per session (limits)

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

Setup guide: install, configure and run RE-call  ·  Validity Frontmatter: the open spec RE-call implements

Why RE-call

Nearest-match retrieval cannot tell the difference between what is true and what merely reads like it. When a corpus keeps its history, and real agent memory does, the retracted claim and its correction are both retrievable, and the retracted one is often the nearer match. That is not a tuning problem. A ranker with no notion of validity has no way to prefer the correction.

RE-call came out of a production, long-running trading-research agent: months of operation, 792 typed memos, 6,469 chunks, re-indexed daily by a session-end hook (counts from the private corpus behind the case study, so no committed artifact backs them). Every guard in this repository exists because that agent failed a specific way without it. See docs/CASE_STUDY.md.

It is for teams putting agent memory behind real applications, where a stale or unsupported memory is worse than no memory: 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. Memory that knows what it no longer believes, and says so.

How that compares to the usual choices (feature rows; the only measured column is Mem0, from the paired head-to-head in benchmarks/REVIEW.md):

	RE-call	Mem0	Zep / hosted memory	Plain pgvector / Chroma
LLM calls to build memory	none	one extraction call per session (measured: 272 calls for the LOCOMO corpus RE-call built at zero)	provider-dependent	none
Runs on your own database	yes, PostgreSQL + pgvector	self-host or SaaS	SaaS first	yes
Supersession and validity	declared in frontmatter, enforced per hit	no equivalent	no equivalent	none
Explicit abstention	calibrated threshold, refusal with a reason	no	no	no, top-k always answers
Trust metadata per hit	verdict, confidence, cosine, provenance, tenant	score	score	score
License	Apache 2.0	Apache 2.0	proprietary SaaS / OSS core	Apache 2.0 / MIT

The rows for Zep and plain vector stores are feature comparisons, not measurements; nothing here claims a benchmark against them.

The vocabulary that carries that validity, supersedes, valid_from and valid_until in a document's frontmatter, is published separately as Validity Frontmatter: MIT licensed, with a zero-dependency TypeScript implementation beside it. RE-call is its Python implementation, not its owner. The specification is deliberately licensed more permissively than this repository, so adopting the vocabulary carries no obligation to adopt the engine.

Capability	What it means in practice
Validity-aware retrieval	Superseded, expired, not-yet-valid, low-confidence, and not-entailed hits are surfaced as verdicts rather than flattened into ordinary search results.
Explicit abstention	When no valid result clears the calibrated threshold, callers receive an abstention with a reason instead of a nearest-neighbor guess.
Local operation	Ingest and retrieval run on PostgreSQL plus pgvector. Local embeddings are supported, so memory can be built and queried without a memory-layer LLM call.
Policy-driven configuration	Embedder, reranker, calibration, trust policy, and retrieval profile are selected to match legal, hardware, latency, quality, and cost requirements. The default is local and offline; higher-quality or hosted options are opt-in.
Production boundaries	Tenant IDs, row-level security, token-scoped MCP HTTP transports, erasure, quotas, timeouts, migrations, and observability are part of the shipped surface.
Reproducible evidence	Published numbers are tied to committed artifacts, and the claim gate checks them in CI.

Measured strengths:

Strength	Evidence boundary
Lower memory-layer cost	The LOCOMO head-to-head records no RE-call memory-layer LLM calls, while the comparator pays for extraction calls. See benchmarks/REVIEW.md.
External abstention check	On MTRAG, IBM's multi-turn RAG benchmark, RE-call is second on correct refusals among the recomputed systems and stays near the top answer-quality rows. See docs/MTRAG_BENCHMARK.md.
Retrieval on a third-party personal-memory benchmark	On ATM-Bench, across 1,013 questions of personal memory QA, the benchmark's own evaluator scores this run at Recall@10 92.8924 and QS 68.4264 . The leaderboard submission is open, not accepted, the answer model is not matched to the published baselines, and both limits are stated in docs/ATM_BENCH.md.
Validity beats nearest-match retrieval	Declared supersession makes the current memory win over stale but similar memory. The larger trust study is in results/FINDINGS.md.
Stronger than a plain vector store	Returned hits carry verdicts, confidence, provenance, tenant scope, and validity metadata. Plain top-k retrieval returns neighbors and leaves trust to the caller.
Clear limits	The evidence states where RE-call works, where it does not, and when a corpus-specific measurement is required.

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.

Quickstart

Two commands, and the second one starts its own database:

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

The distribution is recall-rag; the import and the command are recall. The name recall on PyPI belongs to an unrelated package, so pip install recall gets you something else entirely. Do not install both into the same environment.

That provisions a throwaway PostgreSQL with pgvector in Docker, indexes a small corpus that ships inside the package, and answers three questions: one it can answer, one whose nearest match is a claim that was later retracted, and one it refuses. The middle one is the point.

Measured 2026-08-22 on one Windows machine with the pgvector image already pulled: about 50 seconds cold, and about 22 seconds on a re-run that reuses the container. A machine without the image also pays for that pull, which is the largest and most variable part and is not included here. Re-measure with time recall quickstart.

Nothing is calibrated and nothing is registered with an agent. It prints the next command for each.

recall quickstart --remove          # stops the database and destroys its volume

Already running PostgreSQL with pgvector? recall quickstart --existing-dsn <dsn> skips Docker entirely.

The full install

The quickstart is a demonstration, not an install: it answers questions about a sample corpus with an uncertified threshold, and it leaves your own notes untouched. What follows is the different and longer thing, which points RE-call at your memory, fits a threshold to it, and registers the MCP server with your agent.

RE-call keeps memory in your own PostgreSQL with pgvector, so a database comes first.

Already running PostgreSQL with pgvector? Skip ahead and point the DSN at it.

Want a throwaway one? Save this as docker-compose.yml, then start it:

services:
  db:
    image: pgvector/pgvector:pg18
    environment:
      POSTGRES_USER: recall
      POSTGRES_PASSWORD: recall
      POSTGRES_DB: recall
    volumes:
      - recall_pgdata:/var/lib/postgresql
    ports:
      - "5432:5432"
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U recall"]
      interval: 2s
      timeout: 3s
      retries: 30

volumes:
  recall_pgdata:

docker compose up -d --wait

Then install, create the schema, and run the guided setup wizard. The wizard records the selected embedder, retrieval options, and an optional calibration that is fitted to your labeled queries and your corpus.

pip install "recall-rag[fastembed]"
python -m recall.cli --migration-dsn postgresql://recall:recall@localhost:5432/recall schema --dim 384 apply
python -m recall.cli setup

Those three run unchanged in PowerShell.

The schema command targets the default chunks table deliberately. Global migrations have to be applied there before any other table, so starting with --table something_else on a fresh database stops with SchemaTooOld. To add a separate index later, apply the default target first, then pass --table.

When the wizard asks whether to calibrate, it wants a labeled query file and the corpus those queries refer to. You do not have to build either to try it: both ship inside the installed package, next to each other.

python -c "import recall.eval, pathlib; print(pathlib.Path(recall.eval.__file__).parent)"

That prints a directory holding queries.json, a labeled set covering both answerable and unanswerable questions, and corpus/, the documents those questions are labeled against. Give the wizard those two paths and calibration runs end to end. Sources: recall/eval/queries.json and recall/eval/corpus/.

A calibration fitted that way belongs to that sample, not to your data: it shows the mechanism working and gives you a labeled file to copy the shape of. What makes a calibration valid, when a changed corpus needs a new one (recall calibration drift measures that), and what a labeled file must contain are covered in docs/FIRST_CALIBRATION.md and docs/CALIBRATION.md.

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"]

    TR -. "explicit opt-in" .-> RG["Reasoning graph projection"]
    DB -. "generation-bound" .-> RG
    RG --> IP["Inference proposals: review candidates"]
    TR --> RP["Reasoning policy plus budget"]
    IP --> RP
    RP --> RV{"Citation and trust validation"}
    RV --> ROUT["Cited answer, needs review, clarification, or ABSTAIN"]

Product surface

Area	Ships today
Retrieval	Dense, sparse, hybrid RRF, optional SPLADE, optional cross-encoder reranking, calibrated confidence, provenance, and trust verdicts.
Configuration	Guided setup, local and hosted embedder choices, retrieval cost profiles, optional reranking, strict or development trust policy, and per-corpus calibration.
Storage	PostgreSQL with pgvector, ordered SQL migration path, immutable generations, incremental indexing, pruning, and source-scoped erasure.
Agent integration	CLI, MCP server, LangChain retriever, LlamaIndex retriever, and injectable search seams for tests.
Reasoning	Explicit opt-in reasoning API, CLI, and MCP tools over trusted retrieval, generation-bound graph projections, proposal inspection, budgets, and citation validation.
Security	Tenant isolation, row-level security checks, serving and migration DSNs, bearer-token HTTP transports, scopes, quotas, and unsafe-DSN refusal.
Operations	Timeouts, reconnect policy, structured logging, counters, latency percentiles, and MCP stats.
Quality gates	Real pgvector integration tests, type checking, linting, dependency audit, claim-artifact checks, and regression fixtures for known failure modes.

Deliberately out of scope: an end-user dashboard, entity synthesis, high availability orchestration, automatic truth extraction from prose, and corpus rewrites from inference proposals. Reasoning is opt in, citation constrained, and review aware.

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.

When not to use RE-call

Use something else if you need managed hosting, per-chunk ACLs, 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 with an opt-in reasoning layer, not a general reasoning system. It does not infer every missing supersession edge, prove that an on-topic memory answers a near-miss question, promote proposals into corpus truth, 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.

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. 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:

Topic	Rule
Test database	The test suite drops tables. It uses RECALL_TEST_DSN, never RECALL_DSN.
Default credentials	The MCP server refuses a non-local built-in recall:recall DSN unless RECALL_ALLOW_INSECURE_DSN=1 is set deliberately.
Tenancy	Set RECALL_TENANT or PgVectorStore(tenant=...). Use an unprivileged database role, because PostgreSQL superusers bypass RLS.

MCP

On Claude Code, the plugin does all of this for you, including the hooks and a skill that teaches Claude when to search:

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

It asks for a DSN, a tenant and a trust mode, and keeps the DSN in your OS keychain rather than in settings.json. You still need a database first, which is what recall quickstart above is for. See plugin/README.md.

For every other MCP client, the manual wiring (schema, server block, trust mode) is in docs/USING_WITH_CLAUDE.md. Core tools include recall_search, recall_evidence, recall_index, recall_forget and recall_stats; the authoritative list of all tools is docs/API.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:

Document	Purpose
docs/WRITEUP.md	Architecture and design rationale.
docs/API.md	Supported Python, CLI, and MCP surface.
docs/REPOSITORY_MAP.md	What is product, evidence, benchmark support, and archive.
docs/REASONING_OPERATIONS.md	Opt-in reasoning tools, traces, review policy, and operational behavior.
docs/AUTH.md	Authentication, scopes, and tenant isolation.
docs/MIGRATIONS.md	Migration roles, serving DSNs, and schema operations.
docs/OPERATING_MODES.md	Local, production, quality, hosted, and evaluation deployment modes.
docs/FIRST_CALIBRATION.md	Walkthrough from an indexed folder to a trusted, certified corpus, with the traps named where you hit them.
docs/CALIBRATION.md	Calibration workflow and generation-aware serving.
docs/CASE_STUDY.md	Where the system came from and what is public versus private.
docs/RESEARCH_PROTOCOL.md	How benchmark runs are controlled and audited.

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:

Question	Current evidence
Does declared supersession beat plain similarity search?	Yes, on the authored-edge cases measured in the trust and scale studies.
Can abstention be trusted everywhere?	No. It works on far gaps and fails on near-misses unless a stronger answerability layer is added.
Is retrieval quality universal?	No. Corpus shape dominates, and the measured recommendation is to benchmark your corpus before choosing an embedder.
Is the Mem0 comparison apples-to-apples?	The published head-to-head uses the same LOCOMO questions, generator, judge, and paired tests, with reader-tier limits stated in the benchmark review.
What does MTRAG add?	A third-party multi-turn benchmark with an official judge that gives full credit for correct refusal. RE-call does not top the benchmark, and that boundary is stated in docs/MTRAG_BENCHMARK.md.
What does ATM-Bench add?	A third-party personal-memory QA benchmark over eleven thousand email, image and video items, scored by its own evaluator, where half the questions are graded deterministically rather than by a judge. RE-call's retrieval leads the published board by a wide margin; the answer score is not answer-model-matched and the submission has not been accepted yet. Both limits are stated in docs/ATM_BENCH.md.

Important benchmark documents:

Document	Purpose
results/FINDINGS.md	Interpretation, limits, and negative results.
results/RESULTS.md	Complete result tables.
results/ARTIFACTS.md	Checksum and artifact map for readers auditing a claim.
docs/MTRAG_BENCHMARK.md	MTRAG setup, results, and scope boundaries.
docs/ATM_BENCH.md	ATM-Bench official results, comparability boundaries, and where the remaining loss is.
benchmarks/REVIEW.md	Adversarial review of the LOCOMO comparison.
benchmarks/PREREGISTRATION.md	Pre-registered rules for the main memory benchmark.
benchmarks/archive/preregistrations/README.md	Archived preregistrations for follow-up benchmark arms.

Reproduce

From a git clone (the eval harness is repo-only; it is not shipped in the recall-rag wheel):

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.