Skip to content

Memory & efficiency

Persistent memory across sessions (N2) and token compression. Both are optional interop: Tramalia's core already has file-based memory (N0) and the MCP façade (N1).

The 3 memory tiers

Tier What it is Tool
N0 repo files (AGENTS.md, docs/ai/) + CLI Tramalia's core
N1 Tramalia's MCP façade tramalia mcp
N2 real persistent memory (graph/semantic) Engram · basic-memory · mem0
  • What it is / scope: recall across sessions (decisions, observations) in SQLite, with MCP, CLI, TUI and git-sync.
  • Requires: nothing (binary, Go).
  • Install: brew install gentleman-programming/tap/engram (other OSes: see repo).
  • Tramalia uses it in: doctor detects it; init wires it into .mcp.json if installed; close/handoff --engram export the close (opt-in).
  • Interacts with / rule: opt-in export, never secrets by default. It doesn't replace the repo's evidence; it complements it with cross-session recall.

basic-memory — N2 memory in local Markdown

  • What it is / scope: persistent memory as local Markdown files, via MCP. Fits the repo-first philosophy.
  • Requires: Python (uv).
  • Install: uvx basic-memory · pip install basic-memory.
  • Tramalia uses it in: an alternative to Engram in .mcp.json (not by default).

mem0 — semantic N2 memory

  • What it is / scope: a semantic memory layer for agents.
  • Requires: Python.
  • Install: pip install mem0ai.
  • Tramalia uses it in: an N2 alternative (not by default).

Headroom — token compression / efficiency

  • What it is / scope: compresses tool outputs, logs and context before they reach the LLM (60-95% fewer tokens). Library, proxy, wrapper and MCP modes.
  • Requires: Python (or Node).
  • Install: pip install "headroom-ai[all]" · MCP: headroom mcp install.
  • Tramalia uses it in: only with tramalia init --with-headroom (opt-in; because of its proxy mode, never by default). doctor detects it.
  • Interacts with / HARD MOAT RULE: compression ≠ evidence. The raw output (*-output.txt) and metadata.json are always kept in .tramalia/evidence/. Headroom never modifies, replaces or omits them — it only generates derived views (review-summary.md). Its headroom learn must be redirected to docs/ai/06-intentos-fallidos.md, not write freely.

How they fit together

Engram/basic-memory/mem0 remember across sessions; Headroom makes the context cheaper. None touch the governance core (close, log, evidence pack). Tramalia detects them and wires them opt-in; you decide whether to mount them.