v0.8.2 Open Source Β· MIT OpenClaw Plugin

Memory That Thinks

A 6-layer truth pipeline for AI memory. Every fact is scored, decayed, challenged, and promoted β€” not just stored.

6
Pipeline Layers
36
Memory Columns
<50ms
Recall Latency
∞
Languages

Why Epistemic?

Most AI memory is a flat key-value store. Epistemic treats memory as a knowledge system.

❌ Traditional Memory

  • Store everything, hope for the best
  • No confidence scores
  • Contradictions pile up silently
  • Old facts never fade
  • One channel = one identity silo
  • No audit trail

βœ… Epistemic Memory

  • 6-layer pipeline validates every claim
  • Sigmoid-calibrated confidence scoring
  • Automatic conflict detection & resolution
  • Temporal decay with class-aware half-lives
  • Cross-channel identity unification
  • Full JSONL audit log of every event

Features

Everything an AI needs to remember β€” with the rigor it deserves.

πŸ”¬

6-Layer Truth Pipeline

L0 Policy β†’ L1 Normalize β†’ L2 Score β†’ L3 Conflict β†’ L4 Embed β†’ L5 Route. Every claim validated before storage.

⏳

Temporal Decay Engine

Exponential decay with class-aware half-lives. STABLE persists, EPHEMERAL fades. Access refreshes memory.

πŸ›‘οΈ

Trust Resolution

Channel-aware trust levels. Owner gets full weight, group chat penalized, agent inference cautious.

🌊

Entropy Monitoring

Real-time contradiction density. Pipeline halts when entropy exceeds safety thresholds.

🧠

Episode Distillation

Conversations distilled into episodic summaries for long-term context recall.

πŸ”

Hybrid Search

Vector similarity + keyword search + reranking. Context-aware recall at the right time.

🌐

Cross-Channel Identity

Unified subject across Telegram, Zalo, Discord. One person = one memory graph.

πŸ₯

Proactive Health Alerts

Auto-detects expiring records, unresolved challenges, and decaying facts before data loss.

πŸ”’

Channel Isolation

Prevent group chat facts from leaking into DM. Configurable per-channel with owner override.

πŸ“Š

Confidence History

Full audit trail of every confidence change β€” stores, decays, promotions with timestamps.

πŸ”—

Memory Linking

Automatic claim graph with causal, temporal, and contradictory edges. Memories evolve with evidence.

πŸ“

Full Audit Log

Every store, drop, revoke, inject, and conflict event logged in JSONL for transparency.

Quick Start

From zero to first memory in 3 steps.

1

Install

git clone https://github.com/monasprox/epistemic.git cd epistemic && pnpm install
2

Configure

// openclaw.json β†’ plugins.entries.epistemic { "enabled": true, "config": { "embedding": { "apiKey": "your-api-key", "model": "text-embedding-3-small" }, "ownerId": "your-telegram-id" } }
3

Store your first memory

// The AI agent stores memories automatically from conversation. // Or use the tool directly: User: "I prefer dark theme in all editors" β†’ L0: PASS (behavioral claim) β†’ L1: { subject: "user", predicate: "prefers", object: "dark theme in all editors" } β†’ L2: confidence = 0.741 (source: user_explicit) β†’ L5: tier = WORKING β†’ auto-injected into future prompts
Full Installation Guide β†’

Use Cases

Real-world integrations powering AI assistants across channels.

πŸ’¬

Telegram Bot

Personal AI assistant that remembers your preferences, schedule, and context across thousands of messages. Identity-aware: your Telegram ID auto-maps to the memory owner.

πŸ“±

Zalo Integration

Cross-channel memory sharing between Zalo and Telegram. Same person, same memory graph β€” ownerAliases unify identities automatically.

πŸ€–

Multi-Agent Systems

Shared epistemic memory for agent swarms. Trust resolution ensures agent-inferred facts stay cautious while user-explicit claims get full weight.

The 6-Layer Pipeline

How a raw sentence becomes a verified memory.

L0
Policy FirewallClassifies claim type, drops PII, garbage, and non-storable content
↓
L0.5
Sentence ClassifierFilters greetings, questions, commands, code blocks
↓
L1
Claim NormalizerRegex fast-path + LLM fallback β†’ (subject, predicate, object). Vietnamese + English
↓
L2
Confidence ScorerΟƒ(Ξ±Β·source + Ξ²Β·corroboration βˆ’ Ξ³Β·conflict + δ·kind) β†’ 0–1 confidence
↓
L3
Conflict DetectorVector + keyword search for contradictions. CHALLENGED tier on disputes
↓
L4
Embedding + Storagetext-embedding-3-small (1536d) β†’ LanceDB columnar store (36 fields)
↓
L5
Tier RouterQUARANTINE β†’ CANDIDATE β†’ WORKING β†’ FACT based on confidence thresholds
Learn More β†’