Skip to content

memory-core.md

Latest commit

 

History

History
65 lines (49 loc) · 2.93 KB

memory-core.md

File metadata and controls

65 lines (49 loc) · 2.93 KB

Memory Core

Back to Codebase Guide

Gentle-AI wires Engram into agents; Engram owns the memory store. This page explains the boundary so maintainers do not confuse installer code with memory database code.

Store responsibilities

Responsibility Owner in this repository Actual memory owner
Install or download the engram binary internal/components/engram/install.go, download.go External Engram project/runtime
Add MCP config for agents internal/components/engram/inject.go Agent consumes the config
Run engram setup where supported internal/components/engram/setup.go, CLI runtime Engram CLI implements setup behavior
Document user commands docs/engram.md Engram CLI/MCP implementation
Store sessions, observations, prompts, relations, sync mutations Not implemented here Engram store

Memory entities

Gentle-AI docs and prompt assets refer to these Engram concepts, but their schema is not defined in this repo.

Concept Maintainer meaning
Sessions Work periods that can be summarized and recovered later.
Observations Saved decisions, discoveries, bug fixes, patterns, or artifacts.
Prompts User prompts captured so later saves can attach intent.
Relations Semantic links or conflict judgments between memories.
Sync mutations Export/import changes used by Engram sync workflows.

For command and MCP tool descriptions, link to Engram Commands instead of duplicating an API reference.

Save and retrieve flow

AI agent receives prompt
  |
  v
Gentle-AI-installed prompt tells agent to use Engram MCP tools
  |
  v
Agent calls `engram mcp --tools=agent` via configured MCP entry
  |
  +--> save: mem_save / mem_session_summary / related tools
  +--> retrieve: mem_context / mem_search / mem_get_observation
  |
  v
Engram runtime stores and searches memory outside gentle-ai source

Memory invariants

  • MCP command must be stable: internal/components/engram/inject.go prefers stable command paths and preserves existing absolute paths where needed.
  • Agent setup is capability-based: SetupAgentSlug intentionally returns no setup target for agents that use direct config injection.
  • Prompt assets must stay accurate: Engram instructions in internal/assets/ must match the public behavior documented in docs/engram.md.
  • No schema invention: do not document tables, HTTP routes, dashboard pages, or cloud internals unless source or external Engram docs confirm them.

Contributor checklist

  • Decide whether the change belongs to Gentle-AI wiring or Engram itself.
  • Update MCP injection tests when changing config shape.
  • Keep docs/engram.md as the user-facing command reference.
  • Do not read or modify local .engram/engram.db as part of codebase docs or tests.

Navigation

Previous: Repository map | Next: Interfaces