DEV Community

Alp Allovi
Alp Allovi

Posted on

Introducing Ayeixa Agent Memory Spine: Multi-Tenant Episodic Storage & Replay Compaction

#ai #architecture #typescript #opensource

Introducing Ayeixa Agent Memory Spine: Multi-Tenant Episodic Storage & Replay Compaction

Long-running agentic workflows require access to past conversational turns, error precedents, and decision logs. In multi-tenant environments, storing and retrieving historical context presents two major risks: cross-tenant state leakage and context-window token overflow.

Ayeixa Agent Memory Spine (@ayeixa/agent-memory-spine) is an open-source TypeScript engine providing tenant and session namespaced in-memory episodic storage, memory indexing, and token-budgeted replay compaction.

1. Core Architecture

The system consists of four cooperating components:

  1. SessionIsolationEngine: Enforces tenant/session key isolation (${tenantId}::${sessionId}) and verifies that distinct tenant namespaces share zero episodic records.
  2. EpisodicMemoryStore: Structured in-memory storage for episodic events, tags, timestamps, and importance ratings.
  3. ContextualReplayEngine: Retrieves relevant episodes within an exact token ceiling, formatting them into compact replay prompts.
  4. MemoryCompactor: Applies relevance decay and deduplication to maintain high signal-to-noise ratio within limited context budgets.

2. Implemented Capabilities & Test Verification

Verified with hermetic unit tests:

  • Namespaced Isolation: Zero cross-tenant data overlap verified via verifyIsolation (tests/isolation.test.ts).
  • Contextual Replay: Token-budgeted episode reconstruction (tests/replay.test.ts).
  • Memory Compaction: Deduplication and relevance decay algorithms (tests/compactor.test.ts).

Verification: 3/3 hermetic unit tests passing (0 failures).

3. Local Quick Start 

git clone https://github.com/alpallovy/ayeixa-agent-memory-spine.git
cd ayeixa-agent-memory-spine
npm ci
npm run build
npm test
Enter fullscreen mode Exit fullscreen mode

Usage Example 

import { SessionIsolationEngine, EpisodicMemoryStore, ContextualReplayEngine } from './src';

const store = new EpisodicMemoryStore();
const isolator = new SessionIsolationEngine(store);

isolator.storeEpisode("tenant-101", "session-A", {
  id: "ep-1",
  role: "developer",
  content: "Fixed regression in authentication token verification",
  tags: ["auth", "security"],
  timestamp: new Date().toISOString(),
  importance: 0.95
});

const replay = new ContextualReplayEngine(store);
const context = replay.buildReplayContext("tenant-101", "session-A", { maxTokenBudget: 500 });
console.log("Replay Context Episodes:", context.episodes);
Enter fullscreen mode Exit fullscreen mode

4. Limitations & Contributing

  • Pre-release v0.1.0-alpha. Operates as an in-memory store.
  • Public npm publication is pending.
  • Feedback and PRs are welcome! Check out open good first issue tags on GitHub.

License: MIT

Top comments (0)