LOCAL ALPHA OPENREST APIMCP STDIOPOSTGRES 16DASHBOARDPRIVATE DEPLOY

Lore Context.

The control plane for AI-agent memory, eval, and governance.

Know what every agent remembered, used, and should forget before memory becomes production risk.

Run local alpha → View architecture
$ pnpm seed:demo && pnpm smoke:dashboard
Recall@50.928
Precision@50.814
Stale-hit2.1%
p95 latency142ms
lore › context.ledger live · 14.2 q/s
Memories tracked12,840↑ 312/h
Pending review72 redact · 5 flag
Recall@50.928+0.041
p95 latency142ms+8ms
sourcememoryevidenceused_in_responsestale_scoresensitivityreview_status
claude-code#a4f1api-contractyes lowapproved
cursor#b2c8pricing-noteyes lowapproved
qwen-code#c9d3deploy-fixyes medflagged
hermes#d1a7private-keyyes highredact
fastgpt#f3c1old-schemano medreview
context.compose()
eval 0.94
policy pass
trace 142ms
k 5 / 12
gov · gate
memory#a4f1STALESENSITIVE
postgres://lore_local · audit_logall synced 0.4s ago
01 / PROBLEMThe memory gap

Agents remember. Teams need proof.

Every coding agent, support bot, and copilot is now writing to long-term memory. Almost none of them can show the audit trail behind a single answer.

01
14d agonow

Stale memory

Last week's API contract is still being recalled into this morning's pull request. Nobody flagged it.

02
? doc?user?run?when?

Unknown provenance

An agent quoted a fact. Which document, session, user, and run produced it? Without provenance: a shrug.

03
agent !memory

Unsafe writeback

Agents silently persist hallucinations, PII, and one-off conversation noise into shared long-term memory. Forever.

04
pgvecqdrantweav.milvus

Backend lock-in

Every vector store ships a different schema, a different retrieval API, and quietly captures your data. Migrating costs a month.

02 / SYSTEMPipeline

One context plane. Every agent surface.

MCP clients query through a single composer that passes every retrieval through eval, trace, and governance before it reaches the model. Backed by Postgres and a portable agent-memory adapter.

System diagram · context plane v0.4read path · policy gate · persist + audit
MCP CLIENTSSDK · STDIOCOMPOSERCONTROLPERSISTENCE claude-codecursorqwen-codehermesdify CALLcontext.query()k:5 · ttl:24h · scope:repo COMPOSEcontext.compose() retrieve() k=12rerank() localpolicy() redact 2trace() 142msreturn 5 / 12 evalrecall@5 · precision · MRRtracespan tree · per-source · p95governancepolicy · redact · review Postgresaudit · WAL · pgvector1,284,432 rows agentmemoryqdrant · weaviate · pgvec
01MCP clients

Lore exposes MCP stdio and REST surfaces so agents query, write, and review through the same governance layer.

02context.query

A single typed call your agent can rely on. Returns ranked memories with provenance, freshness, and policy state attached to every row.

03composer

MCP clients query through a single composer that passes every retrieval through eval, trace, and governance before it reaches the model. Backed by Postgres and a portable agent-memory adapter.

04eval / trace / governance

Run the same retrieval evaluation against your seed dataset. Watch recall, precision, and stale-hit rate move as you change retrievers, rerankers, and embedding cuts.

05Postgres + adapter

Memory Interchange Format. Export the entire corpus — embeddings, provenance, policy state — and replay it anywhere. No lock-in.

03 / FEATURESProduct surfaces

Built for operators, not memory hype.

Every surface is inspectable, scriptable, and designed to show proof instead of product theater.

F · 01context.query
› context.query("auth flow", k:5)

Context Query

A single typed call your agent can rely on. Returns ranked memories with provenance, freshness, and policy state attached to every row.

stdio · httpk=1..32ttl-aware
F · 02eval.run
run #248 → #276 · +9.2%

Memory Eval Playground

Replay queries against your seed dataset. Tune retriever, reranker, and freshness cutoff. Compare runs side-by-side. Pin the winner.

recall@5precision@5MRR
F · 03observability
composeretrievererank

Memory Observability

A live trace of every retrieval, every write, every redaction. Drill into any span. Filter by source agent, user, or policy outcome.

span treeper-agentp50/p95/p99
F · 04governance
+ memory#a4f1+ scope:repo- secret redactedapproveredactdeferreject

Governance Review

Every writeback passes a policy gate. Human-in-the-loop review for sensitive scopes. Diff, approve, redact, or reject — with audit.

policy.loreapprove/rejectdiff view
F · 05export.mif
lore12,840 mem⇆qdrantreplay

MIF-like Portability

Memory Interchange Format. Export the entire corpus — embeddings, provenance, policy state — and replay it anywhere. No lock-in.

JSON Lines.lore.tarreplay()
F · 06deploy.local
your networklore-apidashboardpostgres⊘

Private Deployment

Single docker compose. No telemetry, no phone-home, no proprietary embedding endpoint. Run it on your laptop, then on your VPC.

docker composeair-gappedBYO Postgres
04 / ALPHAShipped this week

What is in v0.4 alpha.

Honest list. Built first, polished later. Everything below runs offline, on your machine, on a free Docker Compose.

✓REST API/v1/context, /v1/memory, /v1/eval — typed, documentedv0.4.0
✓MCP stdio SDK transportDrop-in for any MCP-compatible clientmcp 0.6
✓Next DashboardTrace explorer, eval comparison, policy review queuenext 14
✓Postgres incremental persistenceWAL-backed, audit log, idempotent writespg 16
✓Demo dataset4,200 seeded memories across 6 fictional teamspnpm seed:demo
✓Playwright smokeDashboard renders + 8 critical paths coveredpnpm smoke
✓Docker Composelore-api · dashboard · postgres in one filedocker
◐Cloud sync (private)Behind a flag · single-tenant · invite onlyflag
◐Eval autopilotScheduled runs against shadow traffic · in reviewnext

Build manifest

v0.4.0 · 28.04.2026
version
v0.4.0-alpha
commit
release tag
runtime
node 22+ · pg 16
smoke
passing · local gate
bundle
static website · Docker images
memory floor
512 MB
telemetry
off by default
license
Apache 2.0
# first rungit clone github.com/Lore-Context/lore-context
pnpm install && pnpm seed:demo
pnpm test # release gate
05 / EVALLast run · 4m ago

Eval proof report. On your own data.

Run the same retrieval evaluation against your seed dataset. Watch recall, precision, and stale-hit rate move as you change retrievers, rerankers, and embedding cuts.

eval/run-0276seed: lore-demo · 4,200 q · retriever: bge-m3 · reranker: cohere-r3LOCAL · v0.4.0-alpha · smoke passing
Recall@50.928+0.041 vs baseline
Precision@50.814+0.022 vs baseline
MRR0.762+0.018 vs baseline
Stale-hit2.1%-0.6pt vs baseline
p95 latency142ms+8ms vs baseline

Per-stage span timing · last 8 traces

embed14ms
retrieve.pgvec46ms
rerank.cohere32ms
eval.score9ms
policy.gate4ms
compose11ms
trace.flush3ms

Top retrieval sources

cursor86%
claude-code74%
qwen-code52%
hermes41%
dify28%
fastgpt12%
4,200 queries · 84,000 memories↗ rerun · diff vs main
06 / INTEGRATIONSDrop-in MCP

Speak the protocols your agents already use.

Lore exposes MCP stdio and REST surfaces so agents query, write, and review through the same governance layer.

CLClaude Codemcp.stdio · alpha
CUCursormcp.stdio · alpha
QWQwen Codemcp.stdio · alpha
OPOpenClawmcp.stdio · alpha
HEHermesmcp.stdio · alpha
DIDifyrest.v1 · alpha
FAFastGPTrest.v1 · planned
RORoo Coderest.v1 · planned
OPOpenWebUIrest.v1 · planned
LOLocal agentsrest.v1 · planned
07 / STARTStart

Start with a local alpha. Prove memory quality before you scale it.

Four commands. A seeded demo dataset. A Playwright smoke pass that proves the dashboard renders. Bring your own Postgres, or use the bundled one.

Run local alpha →View architecture
$ pnpm install
$ pnpm build
$ pnpm seed:demo
$ pnpm smoke:dashboardverified · no remote assets · smoke passing