Auto Memory

Auto-Memory Extraction

Overview

Extract durable learnings and user preferences from a conversation that will remain useful across future sessions. This runs automatically after each agent turn, using a small/cheap model. Quality over quantity: one precise memory is worth more than five vague ones.

Parameters

• conversation (required): The serialized conversation text to extract from
• existing_memories (optional): Already-stored memories to avoid duplicating

Steps

1. Classify Conversation Content

Scan the conversation and classify each substantive exchange into one of these categories:

Extractable:

• Final decisions (user confirmed or explicitly chose something)
• Corrections (user said "no, do X instead" or "that's wrong")
• Stated preferences ("I prefer X", "always do Y", "don't use Z")
• Discovered facts about the environment, tools, or APIs that were verified
• Patterns that were tested and confirmed working
• Bug fixes with root causes identified

Not extractable:

• Intermediate discussion before a decision was reached
• Options that were considered but rejected
• Transient states ("GitHub is down right now")
• Obvious facts any model would know
• One-off task details ("fix the bug on line 42")
• Anything the user explored but didn't commit to
• Version numbers or update confirmations ("updated X to v1.2.3")
• Heartbeat or check-in status reports ("Heartbeat Feb 19: all clear")
• Benchmark scores or run results ("scored 42/89 = 47.2%")
• Bug sweep summaries without a generalizable root cause ("reviewed X, no bugs found")
• UI/feature implementation details ("button text changed to X", "layout uses 3 columns")
• Task completion status ("task X is complete", "run Y failed")
• Project-specific state that won't inform future decisions

Constraints:

• You MUST only extract from the "extractable" category
• You MUST apply the 30-day test: "Would this change a decision I make 30 days from now?" If no, do not extract it. Most conversations produce zero durable knowledge — that's fine.
• You MUST NOT extract intermediate discussion states as settled facts because conversations explore options before deciding, and capturing exploration as truth produces wrong memories
• You MUST NOT extract transient information (outages, temporary workarounds, "currently broken") because these become stale and misleading
• You MUST prefer the final state of a decision over earlier states because users change their mind during conversations

2. Check Against Existing Memories

Compare each candidate extraction against the existing memories list.

Constraints:

• You MUST NOT extract anything that restates, overlaps with, or is a subset of an existing memory
• You MUST NOT extract a weaker version of something already stored (e.g., don't store "use ripgrep" if "Always use ripgrep instead of grep for searching" already exists)
• You SHOULD flag when a new extraction contradicts an existing memory — extract the new one with updated information, as it represents a more recent decision
• You MUST NOT extract more than 1 item total per conversation — force yourself to pick only the single most valuable extraction, or nothing. Most conversations should produce nothing.

3. Draft Extractions

For each valid candidate, draft a concise memory entry.

Constraints:

• You MUST write each entry as a specific, actionable statement — not a summary of what happened
• You MUST keep entries under 300 characters unless additional context is essential for future usefulness
• You MUST use the final decided form, not the discussion form
  • Bad: "User discussed whether to use the rho tmux config or keep the current one"
  • Good: "Rho tmux config swapped in as ~/.tmux.conf, replacing the nix-configs-based one"
• You MUST NOT use vague language like "the user prefers better approaches" because it provides no actionable guidance
• You SHOULD include the "why" when it's not obvious from the "what"
  • Good: "Use printf '%s' instead of echo for piping to jq — echo adds trailing newline"
  • Bad: "Use printf instead of echo"