Ingest
Feed raw text. EidolonDB extracts what is worth remembering.
Self-managing memory for AI agents
EidolonDB
Self-managing memory for AI agents.
Memory that extracts itself. Evolves itself. Stays useful.
The problem
Vector DBs store whatever you give them. You have to figure out what matters.
Agents forget everything between sessions. Context window isn't memory.
Memory bloat is real. Low-signal storage degrades retrieval quality.
How it works
Evolve
short_term becomes episodic, episodic distills to semantic. Noise decays.
Recall
Hybrid search surfaces the right memories at the right time.
Code
import { EidolonDB } from '@eidolondb/client';
const db = new EidolonDB({ url: 'http://localhost:3000', tenant: 'my-app' });
// Memory that extracts itself
await db.ingest("We use Fastify on port 4000. Jordan leads backend.");
// Recall across sessions
const context = await db.recall("project decisions");
// → ["We're using Fastify on port 4000", "Jordan leads backend"]
// Validate a claim against memory
const result = await db.validate("Jordan leads the backend team");
// → { verdict: "supported", confidence: 0.94, reasoning: "..." }Eval results
16 multi-session scenarios · 3 sessions each · held-out LLM judge scoring recall quality and hallucination rejection.
| System | Recall | Hallucination | Overall |
|---|---|---|---|
| No memory baseline | 22% | 0% | 15.8% |
| RAG baseline * | 93% | 93% | 93.3% |
| EidolonDB | 100% | 100% | 100% |
* RAG baseline uses the same false-premise rejection prompt as EidolonDB for a fair retrieval comparison. Naive RAG without this prompt scores ~65%. EidolonDB's advantage grows significantly over longer session horizons.
Features
Auto-Extraction
LLM extracts structured memories from raw text.
Tier System
short_term / episodic / semantic with automatic lifecycle.
LLM Distillation
Episodic memories condense into lasting semantic knowledge.
Hybrid Search
Vector + recency + importance scoring.
Claim Validation
Check any statement against stored memories. Verdict: supported, contradicted, or unverified.
Multi-Agent RBAC
Share memories across agents with scoped read/write grants.
Conflict Resolution
Automatically detect and surface contradictory memories.
MCP Support
Native Model Context Protocol server. Works with Claude, Cursor, Windsurf.
Full API + SDKs
REST API, TypeScript + Python SDKs, zero dependencies.
Simple, usage-based pricing
Start free. Pay as you grow. No per-seat nonsense.
Get started
npm install @eidolondb/client