reason-state

Governed long-term memory for agents. Token-budgeted context + drift control when facts change (explicit retract() + heal()), with deterministic replay/audit. Works keyless in retrieve-only mode.

Use it like a note-taking logger: add facts → retrieve governed context. If you need to update/retract later, give that fact a stable key (optional).

Proof (fast to verify)

• Update/retraction proof pack (keyless): docs/proof-pack.md (npm run bench:proofpack)
• Recall-style evals (key-gated): docs/benchmarks.md (LoCoMo/LongMemEval runners)

Try it in 60 seconds

npm install reason-state

Drop into any OpenAI-style messages[] loop

import { ReasonState } from "reason-state";
import { injectMemoryContext } from "reason-state/integrations/openai";

const rs = new ReasonState({
  apiKey: process.env.OPENAI_API_KEY, // optional
  model: "gpt-4o-mini",               // optional
  maxTokens: 2500,                    // ~10k chars budget
});

// Zero-friction: just add text (we generate an id)
await rs.add("User hates meetings on Friday");

// Recommended for anything you may update/retract later: provide a stable key.
await rs.add("User hates meetings on Friday", { key: "user:pref:no_friday_meetings", type: "fact" });

const goal = "When should we schedule the retro?";
const messages = [{ role: "user", content: goal }];

// Keyless-safe retrieve-only (no LLM): inject governed context into your messages[]
const { messages: withMemory, stats } = await injectMemoryContext(rs, messages, { goal });
// Pass `withMemory` to your model call.
console.log("ROI stats:", stats); // { estTokens, contextChars, unknownCount, dirtyCount, blockedCount, ... }

// Plan (default): builds context, calls your model, applies patches (requires a key)
const { patches } = await rs.plan(goal);

// Explicit governance (no hidden recompute)
await rs.retract("user:pref:no_friday_meetings"); // retract by stable key (blocked + dirty)
await rs.heal();                           // resolves contradictions + reconciles
await rs.rollback("subtree-root-node-id"); // subtree rollback (optional)

Try it with your own logs (no code)

npx reason-state audit path/to/trace.json
# writes: path/to/trace.reason-state.json (Studio-importable)

Demo (1 command)

npm run demo

CLI demo: ingests a few facts, shows balanced retrieval-only context, and if an API key is present runs one planner turn. Works without keys (planner step skipped).

Why teams adopt it

• Lower cost / smaller prompts: governed, token-budgeted context with measurable ROI (stats).
• Drift control: retract + heal when facts change (no prompt spaghetti).
• Debuggable memory: deterministic replay + explicit governance signals (unknown/dirty/blocked).

How it compares (high level)

Capability	Prompt hacking	mem0/Letta-style memory	reason-state
Drop-in to messages[]	✅	✅	✅
Keyless try-now path	✅	✅	✅
Token-budgeted context	⚠️	✅	✅
Retraction + heal as first-class	❌	⚠️	✅
Deterministic replay/audit	❌	⚠️	✅

Benchmarks

• Benchmarks + methodology: docs/benchmarks.md
• Retraction/update proof pack (recommended): docs/proof-pack.md

Docs (short links)

• Adoption guide: docs/adoption.md
• Audit CLI (npx reason-state audit ...): docs/audit-cli.md
• HTTP service: docs/http-service.md
• Keys/tests matrix: docs/keys-and-tests.md
• Hybrid retrieval (optional vector hook): docs/hybrid-retrieval.md
• Studio: docs/ui-studio.md
• Context builder reference: docs/context-builder.md
• System prompt reference: docs/system-prompt.md

Advanced imports

• Advanced engine: import { ReasonStateAdvanced } from \"reason-state/engine\"