Skip to content

memory.md

Latest commit

 

History

History
53 lines (38 loc) · 2.91 KB

memory.md

File metadata and controls

53 lines (38 loc) · 2.91 KB

Memory Architecture

tars has a four-tier memory system. Each tier has different granularity, persistence, and write patterns. All tiers live in the Obsidian vault (TARS_MEMORY_DIR) as plain markdown, readable and editable outside of tars.

Tiers

Tier Files Granularity Persistence Written by
Semantic Memory.md Distilled facts Permanent User (/remember), /review promotion
Procedural Procedural.md Behavioural rules Permanent /review distillation
Daily YYYY-MM-DD.md Notable events per day Rolling window Auto (tools, compaction, extraction)
Episodic sessions/*.md Full conversation summaries Archived Auto (compaction, session save)

Memory flow

[conversation]
    ↓ tool calls, corrections, compaction events
[daily file] ← append_daily() writes timestamped entries
    ↓ automatic fact extraction on compaction/save
[daily file] ← [extracted] facts appended
    ↓ /review promotes good entries
[Memory.md / Procedural.md] ← permanent memory

Automatic fact extraction

After a conversation compacts or ends, the model extracts discrete reusable facts (preferences, project names, decisions, corrections) and writes them to the daily file tagged as [extracted]. Bad extractions age out naturally. Good ones get promoted to permanent memory via /review. Controlled by TARS_AUTO_EXTRACT (default: enabled). Skips trivial conversations (fewer than 3 user messages).

What goes where

  • Semantic memory — "user prefers dark mode", "project uses Python 3.14", "dog's name is Max"
  • Procedural memory — "route 'remind me' requests to todoist", "use metric units for weather"
  • Daily memory — "10:15 added 'buy milk' to todoist", "14:30 session compacted — discussed deployment", "14:30 [extracted] user is deploying to fly.io"
  • Episodic memory — full session summaries with topics discussed, tools used, decisions made

System prompt assembly

Memory is assembled into the system prompt in layers:

  1. Memory.md and Procedural.md are always loaded (small, high-signal)
  2. Today's daily file loads as <daily-context> (cross-session awareness within the day)
  3. Hybrid search (FTS5 + vector KNN) auto-retrieves relevant context on first message using two-pass packing — anchors for breadth, then windowed expansion for depth, under a token budget

All memory blocks are wrapped with an untrusted-data preface to prevent prompt injection from user-edited vault files.

Feedback loop

  • /w flags a bad response — saves to corrections.md
  • /r flags a good response — saves to rewards.md
  • /review distills corrections into procedural rules and promotes good extractions to permanent memory
  • /tidy cleans up stale or duplicate memory entries
  • Procedural rules feed back into every future response via the system prompt