credence

credence

MOST AGENTS REMEMBER TEXT. THIS ONE REMEMBERS EVIDENCE.

credence is an evidence-first agent memory and durable mission runtime. Not a vector store, not a RAG pipeline, not an orchestrator — the layer underneath all of them: an append-only knowledge ledger where every belief carries its evidence, every correction preserves its history, every learning passes an authority gate, and every mission survives its process.

• You need it if your agent runs for hours or days, and "remembering" currently means stuffing text into a context window.
• It's different because knowledge here has epistemics: graded support, evidence links, supersession chains, and an explicit UNKNOWN state. An agent powered by credence can answer "I don't know — and here's the closest thing I do know, and why it doesn't qualify."

KILL THE AGENT. THE MISSION CONTINUES.

(A timed replay of a real npm run demo:kill run — those are actual captured output lines and PIDs, not a mockup. Run it yourself below.)

Run it in 30 seconds

Requires Node ≥ 22.6. No install step, no API key, no model — the whole demo is deterministic and local.

git clone https://github.com/MajidAsghariTabrizi/credence
cd credence
npm run demo:kill

That kill-and-resume is one of four signature moments. Or run them one at a time:

Or run everything: npm run demo.

Every headline claim in this README has a test, a command, or a fixture behind it — see PROOF, NOT PROMISES.

The five ideas this repo owns

1. Memory with epistemics

A claim is not a string in a vector. It is a record:

claim c_e21f0106
  text        "star tracker drifts 0.02°/sol after firmware 2.1 recalibration"
  support     measured            (measured > inferred > reported)
  basis       "post-firmware-2.1 calibration run GC-CAL-0051"
  evidence    e_cal0051 — 21-sol drift series, slope 0.019, r²=0.99
  state       active              (active | superseded | disputed)
  supersededBy []                 (grows — never rewrites)

2. UNKNOWN is a feature

ask returns unknown — with the reason and the nearest rejected candidate — whenever coverage is below threshold or evidence is absent. Retrieval rule → src/kernel/ask.ts · Demo → npm run demo:unknown

3. History is immutable

Corrections supersede. The old claim stays readable forever, linked to its replacement, with the contradiction event between them. Fold → src/kernel/store.ts · Demo → npm run demo:contradiction

4. Missions outlive processes

Mission state is an append-only log on disk (keel). Kill the process; a fresh one folds the log and continues. Mission state is operational state — it is never automatically trusted knowledge. src/kernel/keel.ts · Demo → npm run demo:kill

5. Learning has authority

Agents may discover and propose. Only owner/lead commit durable knowledge. A denied commit is logged, not lost — the proposal waits for review. src/kernel/learn.ts · Demo → npm run demo:governed

PROOF, NOT PROMISES

Benchmarks are synthetic, seeded (deterministic), and honest — latest results on this exact kernel: unknown-detection 1.00, hallucination-on-unanswerable 0, paraphrase recall 0.65 (token-coverage retrieval, no embeddings), ask latency p95 ~1.3 ms at 200 claims. Weak numbers stay published.

Architecture

Agents      worker · lead · watcher · operator      (propose, never commit)
Harness     lifecycle · tools · keel missions       (durable operational state)
Brain       ask · explain · investigate · learn · pulse
Packs       ground-control/ … each an events.jsonl  (append-only, one fold away)

One process, one directory, zero dependencies. The event log is the database — no migrations, no server. Deep dive → docs/ARCHITECTURE.md.

Deep rabbit holes

What happens when the Brain doesn't know?

ask scores every active claim by query coverage (fraction of your question's content tokens present in the claim) weighted by support grade. Below 0.45 — or any best match with zero evidence — you get state: unknown with why and the nearest rejected candidate. You can then investigate: collectors gather findings, split into durable candidates (commit via learn) and ephemeral observations (never commit — a measured value is wrong tomorrow).

What happens when new evidence contradicts old evidence?

A supersede delta writes two events: contradiction-raised (both ids + the note) and claim-superseded (old → new). The fold marks the old claim superseded and appends the successor id. explain <claimId> walks the whole chain. Nothing is deleted; the log only grows.

What exactly is allowed to learn?

Caller identity is a role: owner, lead:*, agent:*, watcher. propose() records the delta as a pending proposal; agent/watcher commits are refused with a learning-denied event — the denial itself is part of history. owner/lead approve with commitProposal. Private claims (sensitivity: private) are filtered out of agent reads at the store layer.

What survives a dead process?

Everything that was appended before the kill: claims, evidence, proposals, mission steps. The keel writes each step after it completes, so a kill mid-step replays that step — never skips, never duplicates. mission-resumed events count interruptions.

Honest status

This is v0.1.0: a small, complete, honest kernel — not a platform. The gaps above are the roadmap.

DeepSeek Harness integration

Credence is an integration target, not a dependency. The adapter exposes ask / learn / pulse to a DeepSeek Harness agent session via the kernel's programmatic API — an agent can consult the ledger and propose learnings (it will be correctly DENIED if it tries to self-commit). Unaffiliated with DeepSeek; the harness is just a good host.

Contributing

Issues and PRs welcome — good first issues are marked. Read CONTRIBUTING.md. The kernel is ~700 lines of dependency-free TypeScript on purpose: read it in one sitting before extending it.

License

MIT · © 2026 Majid Asghari Tabrizi. The synthetic demo domain (probe MNEMOSYNE-7) is fiction; any resemblance to real spacecraft telemetry is a calibration error.