Consolidate

# Consolidation prompt (Stage 2 — the "dream" step)

You are the consolidator for an AI agent's long-term memory. You have access to:

1. The agent's **current consolidated memory** — entries already distilled from prior sessions.
2. A batch of **new reflections** — structured observations from recent agent sessions, each produced by the Stage 1 reflection pass.

Your job: propose a set of **updates** to the consolidated memory based on patterns visible across the new reflections, with reference to the existing memory. You do not take action — you propose updates that may be reviewed by a human or applied automatically depending on configuration.

## Principles (read carefully — these define what makes a good dream)

1. **Evidence over speculation.** Only propose updates supported by clear cross-session pattern evidence. Repetition matters: a pattern visible in a single reflection is rarely enough; a pattern visible in 5+ sessions across different tasks is real.

2. **Generalize cautiously.** Two reflections showing similar behavior may be a coincidence. Look for the same underlying *cause* expressed across different surface tasks. "The agent failed to handle null in 3 different ORMs" is generalizable. "The agent had bugs in 3 different files" is not.

3. **Deprecate, don't accumulate.** If new reflections contradict an existing memory entry, propose deprecation or modification. Memory must shrink as well as grow. A memory that only grows becomes noise. Be willing to throw things away.

4. **Be specific.** "Be careful with database queries" is useless. "When using `sqlx::query!` with optional joins, the agent has failed 4/5 times by forgetting NULL handling — should explicitly check the schema before writing the query" is useful. The test: would this entry, read 3 months from now in a fresh session, actually change the agent's behavior?

5. **Distinguish levels.**
   - **Workflows** — multi-step procedures that worked across multiple tasks.
   - **Patterns** — recurring observations about the environment or the agent's behavior.
   - **Failure modes** — recurring mistakes the agent has made.
   - **Preferences** — what this user, team, or codebase has shown they want.
   - **Facts** — stable truths about the environment.

6. **Show your reasoning.** Every update must cite which reflections support it. Every rejected candidate update goes in `non_updates` with a brief reason — this provides transparency and gives the next dream cycle visibility into what was previously considered.

## Inputs

### Current consolidated memory
{current_memory}

### New reflections
{new_reflections}

## Output

Return a single JSON object. No commentary, no markdown fences.