blackms/aistack/memory

aistack: Shared agent memory

aistack gives agents a persistent, searchable memory store (SQLite-backed, optionally vector-indexed). This is how specialized agents hand off context without stuffing everything into one prompt, and how knowledge survives across sessions.

When to use

• You established a fact, decision, or artifact that a later agent or session will need.
• The user asks to "remember this", "recall what we decided", or "what's the context on X".
• You are about to re-derive something that was likely already computed earlier — search memory first.

MCP tools (bundled aistack server)

List the exact tool schemas with npx @blackms/aistack mcp tools.

Workflow

Before starting heavy work — recall

Call memory_search with the key terms of the task. If a relevant decision or artifact exists, reuse it instead of redoing the work.

After establishing something durable — store

Call memory_store for:

• Architectural decisions and their rationale ("decided to use X because Y").
• Stable facts about the codebase (entry points, conventions, gotchas).
• Hand-off artifacts between agents (an API contract, a test plan, a research summary).

Write the content so a future agent with no context can act on it: be explicit, include file paths, avoid pronouns referring to the current conversation.

Do NOT store

• Secrets, tokens, credentials.
• Large blobs better kept in files.
• Transient scratch state that will not matter next session.

Tips

• Tag memories consistently (e.g. by component or issue id) so memory_search is precise.
• When orchestrating (see the aistack-orchestrate skill), have each agent store its output and the next agent recall it — that is the core handoff mechanism.

Reference: https://github.com/blackms/aistack