loreweave

Most knowledge tools are write-only. You capture diligently, the vault grows, and six months later you can't find the thing you know you wrote — because retrieval is keyword search over prose, nothing ever resurfaces on its own, and nothing notices when what you wrote last year stopped being true.

Loreweave is the layer that fixes that. Point it at a folder of markdown (Obsidian or plain) and it builds a knowledge graph, a bitemporal fact store, and a memory model over your notes — then hands them to you through a CLI and to your AI agents through MCP.

Three guarantees, enforced by the code rather than promised:

• Your files win. User markdown is never mutated; the engine only appends, and only under lore/. The vault is the source of truth — the index is a cache you can delete at any time and rebuild identically (there's a test for that).

• No LLM anywhere in the core. Indexing and retrieval use zero tokens and make zero network calls. Same vault, same query, same answer — forever.

• Memory you can read. Every fact an agent stores is a markdown line you can open, edit, and git diff.

Quickstart

cd ~/my-vault
npx loreweave init          # creates .lore/
npx loreweave index         # incremental sync; one changed note: 32 ms at 1k notes, 656 ms at 20k (see Scale)

npx loreweave search "why did we drop the queue design"
npx loreweave ask "what's the status of project atlas"
npx loreweave dream         # what's duplicated, contradicted, stale, unlinked

Zero configuration and no network: out of the box it runs on BM25 + knowledge-graph spreading activation. Add local embeddings when you want them:

// .lore/config.json
{ "embedding": { "provider": "ollama", "model": "mxbai-embed-large" } }

Pick mxbai-embed-large (670 MB) for quality or nomic-embed-text (274 MB) when disk and indexing speed matter more — required task prefixes are applied automatically for both, and for the E5, BGE and Arctic families. The measured difference between the two is in Benchmarks. No embedding provider means lexical + graph retrieval, still fully functional.

Works with non-English vaults: Chinese, Japanese and Korean text is segmented per character so it is searchable at all, and other scripts index as written.

Related MCP server: Engram

What it does

1. Knowledge that has a timeline. Facts are bitemporal: when they were true in the world (valid_from/valid_until) and when the system learned them (recorded_at). Contradictions supersede rather than overwrite, so history stays queryable.

$ lore assert "Ledger Format" status draft --valid-from 2026-01-01
$ lore assert "Ledger Format" status final --valid-from 2026-08-01
✓ Ledger Format :: status :: final
  superseded: "draft" (now valid until 2026-08-01)
  journal: lore/journal/2026-08-01.md

(lore and loreweave are the same binary — npm i -g loreweave gives you both; npx loreweave works without installing.)

Both time axes are queryable, which is what makes it bitemporal rather than merely historical. --as-of asks what was true then; --as-known-at asks what was believed then. They disagree exactly when you learn something after the fact — which is when you most need to reconstruct what a past decision was actually based on:

$ lore facts --subject Vendor --as-of 2024-06-01
Vendor :: reliability :: poor — outage postmortem  (2024-01-01 → now)

$ lore facts --subject Vendor --as-known-at 2024-06-01
Vendor :: reliability :: good  (2024-01-01 → 2024-01-01)  [superseded]

Which fact wins is decided deterministically (newest valid-time, provenance as tiebreak) — never by asking a language model which one looks fresher. And the whole history of anything is one command — every value change merged chronologically with the dated prose that mentions it:

$ lore timeline Project Atlas
2024-01-15  status: planning  (until 2024-09-01)
2024-02-10  • [[Project Atlas]] kicked off with a three-person crew.  [kickoff.md]
2024-09-01  status: planning → active
2025-06-20  • The [[Project Atlas]] midpoint review went long but well.  [review.md]

Temporal-graph systems build this by running an LLM over every document at ingestion. Here the supersede chain has been maintained all along, so it is a read-side join: no LLM, no network, same answer every time.

2. Retrieval that follows connections, not just words. Queries fuse BM25, dense similarity (when configured), and Personalized PageRank over the vault's own graph — wiki-links, shared entities, tags, co-occurrence. Two-hop neighbors surface even when they share no vocabulary with your query, and every result tells you why:

• data/glacier-dataset.md#@0  (0.0327)  ⟨via amara osei⟩
  The Glacier Dataset holds meltwater sensor readings from 2019-2024.

3. Memory with dynamics. Every passage carries FSRS-style stability and retrievability — a power-law forgetting curve. Passages that actually get used (not merely retrieved) decay slower; important-but-fading knowledge gets surfaced for review instead of silently rotting. Nothing is ever deleted.

4. It dreams. lore dream is an idle-time consolidation pass that reviews the vault and reports duplicate passages, contradicted facts, stale knowledge, missing links between notes that clearly belong together, and orphans. With --apply it writes a digest and a review queue — append-only, under lore/. It never rewrites your prose: LLM-driven whole-file rewriting is a documented failure mode — each rewrite quietly drops details until the file collapses to mush — so the architecture forbids it.

5. Questions retrieval can't answer. Counting, grouping, and date-range queries run as deterministic SQL over the fact store, not as vibes over embeddings:

$ lore count --predicate trip_to --since 2025-01-01 --until 2025-12-31
    2  Japan
    1  Kenya

6. Facts come from your notes, not from a form. Nobody hand-writes - [fact] X :: y :: z, so the extractor mines the conventions vaults already use:

status: shipped              # frontmatter
- owner:: Priya              # Dataview inline field
- [location] Hyderabad       # Basic Memory observation

Only unambiguous field syntax is accepted automatically. Prose formatting like - **Owner:** Priya is precise on entity notes and noisy on report notes, so it is opt-in (facts.extract: "all") — or an agent can review candidates via lore_propose_facts and assert the real ones. Judgement stays out of the index.

7. Time means when it happened, not when you saved the file. --since and --until filter on content time — frontmatter dates, dated filenames (2025-03-14-standup.md), or dates in the text — falling back to file mtime only when a note carries no date of its own. lore watch keeps the index current so you never have to remember to reindex.

8. Built for agents. An MCP server exposes 15 typed tools so Claude Code, Cursor, or any MCP client can use your vault as durable memory. Session continuity is a query, not a paraphrase — lore resume returns exactly what changed since the agent last connected, computed from record time:

$ lore resume
since 2026-08-11 15:55
~ lore/journal/2026-08-11.md
+ Project Atlas :: status :: shipped (since 2026-08-11)
± Project Atlas :: status: active → shipped

$ lore resume
since 2026-08-11 15:56
nothing changed

Alternatives that summarize the previous session with an LLM inject a paraphrase; this is a deterministic diff. Full setup in Agent memory.

The CLI

Use it as agent memory (MCP)

// Claude Code: .mcp.json  (or claude_desktop_config.json)
{
  "mcpServers": {
    "loreweave": {
      "command": "npx",
      "args": ["-y", "loreweave", "--vault", "/path/to/vault", "serve", "--mcp"]
    }
  }
}

Facts asserted through MCP are written back to lore/journal/YYYY-MM-DD.md as readable markdown, so an agent's memory is something you can open, edit, and git diff:

- [fact] Ledger Format :: status :: final {valid_from=2026-08-01, confidence=0.9, source=stated}

Delete .lore/ and reindex — every fact and edge is reconstructed from those files.

Compared to hosted memory services (Mem0, Zep): those run LLMs at write time to extract and summarize into their own store; loreweave runs no model in the core, keeps memory in your files under your version control, and makes every retrieval reproducible. The trade: they do abstractive summarization, this engine deliberately does not. Retrieval quality against their published benchmarks is below — with the caveats stated, because most published agent-memory numbers measure end-to-end QA with an LLM, which is a different quantity than retrieval.

Benchmarks

These are third-party benchmarks with relevance labels nobody here chose. All loreweave numbers are retrieval metrics — it finds the evidence, it does not write the answer — so they are not comparable to the end-to-end QA accuracy quoted by systems that put a language model after retrieval. Reproduce any number: docs/benchmarks.md. Field comparison with the traps called out: docs/scoreboard.md.

LongMemEval_S (ICLR 2025) — 500 questions, each with ~50 sessions of chat history; find the sessions holding the evidence. Session-level recall, all 500 questions:

A third-party benchmark of the same task reports BM25 alone at 86.2% R@5, BM25+vector hybrid at 95.2%, and a vector-only system at 96.6%. One caveat before the comparison: their metric scores a question 1 if any gold session is retrieved, ours scores the fraction of gold sessions found — ours is the stricter definition, so treat cross-system gaps of under a point as noise. With that stated: at 95.9% loreweave is past the hybrid and just short of the vector-only system, with a local model, no network, and a lexical index it can fall back to.

BEIR / SciFact — 5 183 scientific abstracts, 300 claims, scored by nDCG@10 (ranking quality: rewards putting relevant documents nearer the top):

LoCoMo — 10 long conversations, 1 982 evidence-labelled questions, turn-level recall. No comparable third-party retrieval number exists (published LoCoMo results are LLM QA accuracy), so these are offered as a target rather than a comparison:

What the numbers say, plainly. Recall is strong and rank-1 is the weakness — LongMemEval R@10 0.983 vs R@1 0.590 — which is why figures here are quoted at R@5 and why an agent consuming these results should read a top-5 list, not trust rank 1. The single biggest quality lever measured is the embedding model itself: mxbai over nomic is worth more than every downstream tuning combined on document (BEIR) and session (LongMemEval) retrieval — and measures flat on LoCoMo's single-turn passages, so it is a scale effect, not magic. A cross-encoder reranker is available (rerank in config) but earns its keep only for rank-1 consumers without embeddings — stacked on embeddings it loses recall on both public benchmarks, so leave it off unless that trade is yours. Full analysis, including the failed experiments and the defect that benchmarking caught: docs/evaluation.md.

Loreweave also ships an internal regression benchmark (npm run eval) over three purpose-built vaults — including a temporal-perturbation test where the shipped config scores 100% consistency vs BM25's 0%, gated in CI on every push. Internal corpora are good for regression and worthless as proof, so the details live in docs/evaluation.md rather than here.

Scale

Measured, like the quality numbers — npm run scale reproduces this on your own machine (synthetic vaults, 3 blocks per note, dense interlinking):

Full index scales at 0.93× per note from 5 k to 20 k — linear or better, no superlinear step hiding in the middle. "Incremental" is one changed note, which is what lore watch actually does all day. Everything here is one process, one SQLite file, no daemon — and search at 3-38 ms p50 is fast enough to sit inside an agent loop.

How it works

vault/*.md ──parse──▶ notes · blocks · wiki-links · tags · entities
                            │  (incremental: mtime + content hash)
                            ▼
                  SQLite .lore/index.db  ── disposable cache, rebuildable
                            │
        ┌───────────────────┼────────────────────┐
        ▼                   ▼                    ▼
   graph (CSR)         retrieval             facts
   blocks ∪ entities   BM25 + dense + PPR    bitemporal, supersession,
   2-iteration PPR     → weighted RRF        deterministic freshness,
   α = 0.5             → FSRS boosts         aggregates
        └─────────┬─────────┴──────────┬─────────┘
                  ▼                    ▼
             dream (idle-time)     CLI · MCP

Design rules the code enforces:

• Files win. User markdown is never mutated. The engine only appends, and only under lore/.

• Invariants in code, not prompts. Schema, migrations, graph construction, and supersession are typed, versioned, and tested — no LLM re-specifies them at runtime.

• No LLM required anywhere in the core. Indexing and retrieval use zero tokens. Language models are consumers of this engine, not dependencies of it.

• Everything is re-derivable. A full rebuild reproduces byte-identical derived state (there's a test for that).

Research lineage

Every significant choice traces to 2024-2026 literature; the survey lives in docs/research/.

Library use

import { openContext, indexVault, search, assertFact, queryFacts, dream } from 'loreweave';

const ctx = openContext('/path/to/vault');
await indexVault(ctx.store, ctx.root);
const hits = await search(ctx, 'streaming compaction', { k: 5 });
assertFact(ctx, { subject: 'Atlas', predicate: 'status', object: 'shipped', validFrom: '2026-08-01' });
const asOfMarch = queryFacts(ctx.store, { subject: 'Atlas', asOf: '2026-03-01' });
const report = dream(ctx);
ctx.close();

Development

npm install
npm test          # 440 tests
npm run eval      # retrieval benchmark vs BM25 baseline
npm run typecheck
npm run build

Requires Node ≥ 20. Single native dependency (better-sqlite3). Tested in CI on Linux, macOS and Windows across Node 20 and 22.

License

MIT © Ambuj Upadhyay