@adaptivemcp/memory

Persistent operational knowledge for Adaptive MCP, backed by SQLite (node:sqlite).

memory is the store. Every other package reads from or writes to it: telemetry folds execution events in, evaluation writes derived insights back, and the extension derives its YAML view from it. The YAML is a projection, never the source.

Install

npm i @adaptivemcp/memory

Requires Node 22+ (Node 26 recommended; the node:sqlite module is available without the --experimental-sqlite flag).

Usage

import { MemoryStore } from "@adaptivemcp/memory";

// :memory: for ephemeral; pass a file path to persist.
const store = new MemoryStore({ path: ":memory:" });

store.ensureTool("deploy_service", "ci");

// Record an observation (typically done by @adaptivemcp/telemetry).
store.recordExecution("deploy_service", {
  status: "completed",
  durationMs: 1200,
  model: "gpt-5-mini",
  cost: { amount: 0.0021, currency: "USD" },
});

// Static, human-authored metadata.
store.setAnnotation("deploy_service", { risk: "high", note: "Needs approval." });

// Derived, learned metadata (typically written by @adaptivemcp/evaluation).
store.addInsight("deploy_service", {
  key: "observed_failure_rate",
  value: 0.12,
  confidence: 0.9,
  source: "evaluation",
});

store.close();

API

| Method | Purpose | | --- | --- | | ensureTool(toolName, serverName?) | Create a tool record if absent. | | getTool(toolName) | Read a single ToolRecord. | | allTools() | Read every ToolRecord. | | setAnnotation(toolName, annotation) | Write a static Annotation (risk: low|medium|high). | | addInsight(toolName, insight) | Append a learned Insight. | | addRecommendation(toolName, recommendation) | Append an advised Recommendation. | | clearRecommendations(toolName) | Drop stale recommendations before re-planning. | | recordExecution(toolName, partial) | Fold a completed/failed execution into ToolStats. | | close() | Close the underlying SQLite database. |

The store schema is a single tools table (tool_name PK) holding the annotation, insights, recommendations, and stats as JSON columns.

License

Released under the MIT License. Copyright (c) 2026 Kemal Elmizan.