thoughtbox:evolution

Thought Evolution (A-Mem Pattern)

When you add a new insight to a reasoning session, earlier thoughts don't automatically update. Thought 1 might say "consider rate limiting" while thought 15 decides "use sliding window algorithm" — but thought 1 doesn't know about the sliding window decision. This skill checks which prior thoughts should evolve.

Based on the A-Mem paper (arxiv.org/abs/2502.12110): when new memory is added, find related existing memories and update their context.

When to Trigger

Run evolution checks when the new thought:

• Resolves ambiguity from earlier thoughts
• Contradicts an earlier assumption
• Adds implementation detail to a high-level earlier thought
• Synthesizes multiple earlier threads into a conclusion

Don't run for every thought — only on significant ones (synthesis, conclusions, decisions, revisions).

Workflow

Phase 1: Retrieve Session Content

// thoughtbox_execute
async () => {
  const session = await tb.session.get("current-session-id");
  return session.thoughts.map((t, i) => ({
    number: t.thoughtNumber,
    content: t.thought.slice(0, 200)  // Truncate for efficiency
  }));
}

Phase 2: Spawn Evolution Checker

Dispatch a Haiku subagent for cost efficiency (~400 tokens in subagent context, ~50 tokens returned):

Spawn subagent (model: haiku):
  "Evaluate which prior thoughts should be updated based on a new insight.

   NEW INSIGHT:
   [Your new thought content]

   PRIOR THOUGHTS:
   S1: [thought 1 content]
   S2: [thought 2 content]
   ...

   For each thought, respond ONLY with:
   S1: [UPDATE|NO_UPDATE] - [brief reason if UPDATE]
   S2: [UPDATE|NO_UPDATE] - [brief reason if UPDATE]
   ...

   Be selective. Only suggest UPDATE if the new insight meaningfully enriches
   the prior thought's context. Keyword overlap alone is not enough."

Phase 3: Apply Revisions

For each thought marked UPDATE, create a revision:

async () => {
  await tb.thought({
    thought: "EVOLVED: [original content] — Now contextualized: [how new insight relates]",
    thoughtType: "reasoning",
    isRevision: true,
    revisesThought: 1,  // The thought number being updated
    thoughtNumber: 20,  // Current thought number (advances the chain)
    totalThoughts: 25,
    nextThoughtNeeded: true
  });
}