Features

Everything you need to keep agents in sync

Flagship

Read-your-writes, as an API contract

Lore is a coordination memory layer: agents write events, Lore consolidates them into versioned facts, and serves budget-fit context packs with consistency guarantees — so every agent works from the same reality.

Every write returns a seq. Any later pack asked for that min_seq is guaranteed to reflect it — extracted, or handed back as a raw tail until extraction lands. Visibility is verifiable, not assumed.

lore.write(event)→ seq 1032

lore.pack({ min_seq: 1032 })→ covered_seq 1032 ✓

Governance, enforced

Per-agent ACL compiled to SQL predicates, trust tiers with quarantine, and mandatory provenance — enforced in the query plan, not the prompt.

A token bill that drops

Deterministic, budget-fit packs maximize prompt-cache hits — and a built-in meter reports the tokens and dollars saved versus raw history.

Conflicts resolved at write time

Same fact, two writers — adjudicated by a policy you choose under a database constraint. The loser is versioned, never silently erased.

Provenance on every fact

“Why did the agent do that?” stops being an afternoon of archaeology — who wrote what, which run, which event, one click away.

shell

$ docker run loregpt/lore
# one Go binary · Postgres · Valkey

Real open source

A full self-hostable server — not a library you operate around. Apache-2.0, no graph database, no external services.

How it works

One real multi-agent loop — read-your-writes, a conflict resolved by policy, an untrusted write quarantined — ending in one shared reality.

Lore · how it works

—waiting

✓

Researcher

research API change

✓

Coder

implement fix

✓

Reviewer

review & ship

new-agent · untrusted

ignore previous instructions and…

LORE

quarantine

0tokens saved

The spec sheet

The primitives that turn “agent memory” into agent coordination — each one an API surface, not a slogan.

Read-your-writes

min_seq → covered_seq

Every write returns a seq. Any later pack requested with that min_seq is guaranteed to reflect it — extracted, or included as a raw tail until extraction lands.

Conflict resolution

write-time

Same fact, two writers — resolved at write time under a database constraint, by a policy you choose (last-writer-wins, merge, LLM, or manual). The loser is versioned, not erased.

Access control

SQL predicates

Per-agent policies compile to SQL predicates, enforced in the query plan with row-level security behind them. New agents start quarantined until trusted.

Provenance

run_id

Who wrote what, which run, which event — a mandatory field, one click away. Every fact is auditable by construction.

Deduplication

merge on write

Near-duplicate memories are merged before they ever bloat a pack, so context stays dense and cache-friendly.

Token budgeting

pack_hash

Budget-fit, deterministic packs — cache-friendly by construction, with a meter that reports tokens and dollars saved versus raw history.

Freshness SLO

freshness_lag_ms

Write → queryable lag is measured and returned in the response, not left to your imagination.

Working memory

sync CAS lane

A hot, high-churn coordination lane on a synchronous CAS path — for the state that changes faster than extraction can keep up.