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"

4. Categorize

Assign each extraction a type and category.

Learning types:

• Corrections, discovered patterns, environment facts, bug root causes, tool behaviors

Preference types with categories:

• Communication: Voice, tone, formatting, emoji policy
• Code: Style, patterns, conventions, architecture
• Tools: Tool preferences, CLI flags, configuration
• Workflow: Process, git, deployment, review practices
• General: Anything that doesn't fit above

Constraints:

• You MUST assign exactly one category per preference
• You SHOULD default to "General" only when no other category fits

Output Format

Output strict JSON only:

{
  "learnings": [
    {"text": "concise, actionable learning statement"}
  ],
  "preferences": [
    {"category": "Communication|Code|Tools|Workflow|General", "text": "concise, actionable preference statement"}
  ]
}

If there are no genuinely new, durable items to extract, return:

{"learnings": [], "preferences": []}

Returning empty is better than returning noise.

Examples

Example 1: Decision After Discussion

Conversation excerpt:

User: "Should we source the rho tmux config or swap it in?" Agent: "Here are the differences... Want me to source it or swap it?" User: "swap it in" Agent: swaps the config

Good extraction:

{"learnings": [{"text": "Rho tmux config at ~/.rho/tmux.conf is now the active ~/.tmux.conf, replacing the nix-configs-based one. Old config backed up at ~/.tmux.conf.bak."}]}

Bad extraction (captures discussion, not decision):

{"learnings": [{"text": "The rho-specific tmux configuration is not being used by the current ~/.tmux.conf"}]}

Example 2: User Correction

Conversation excerpt:

Agent: drafts a formal, structured X post User: "do something more clever" Agent: drafts dry, self-aware version User: "I like that personality, do that more"

Good extraction:

{"preferences": [{"category": "Communication", "text": "X post voice: understated, self-aware, dry. State the problem, state what was done, land it flat. Let the reader connect the dots."}]}

Bad extraction (too vague):

{"preferences": [{"category": "Communication", "text": "User prefers clever X posts over formal ones"}]}

Example 3: Nothing Worth Extracting

Conversation excerpt:

User: "find recent X posts about rho and post something" Agent: searches, drafts, posts

Correct output:

{"learnings": [], "preferences": []}

The task was executed but no durable knowledge was produced.

Example 4: Tempting But Not Durable

These look like learnings but fail the 30-day test:

Candidate	Why it fails
"pi-coding-agent updated to 0.55.4 on 2026-03-04"	Version snapshot — stale tomorrow
"Heartbeat Feb 19 08:02 UTC: rho-web healthy"	Status report — not a decision or pattern
"ChefBench scored 42/89 = 47.2% on Terminal-Bench 2.0"	Benchmark result — won't change future behavior
"Fresh-eyes bug sweep: reviewed X, no bugs found"	Sweep status — no generalizable root cause
"The sessions hamburger button should use an icon-only button"	UI detail — too specific to one feature
"All 224 unit tests pass with no failures"	Test status — transient fact
"Task X is complete and verified"	Completion status — belongs in task tracking, not memory

Correct output for all of the above:

{"learnings": [], "preferences": []}

Troubleshooting

Memory Growing Too Large

• Return empty rather than extracting marginal items
• Prefer updating/superseding existing memories over adding new similar ones

Contradicts Existing Memory

• Extract the newer version — it represents a more recent decision
• The memory system handles dedup and supersession separately