AutoDreamMemory Consolidation

Over many conversations, your memory files accumulate duplicates, stale facts, and contradictions. AutoDream uses the AI itself to clean this up: reading every memory file, merging duplicates, removing outdated entries, and keeping your MEMORY.md index concise and actionable.

• Deduplicates near-identical entries across topic files
• Removes facts contradicted by newer information
• Converts relative dates ("yesterday") to absolute (YYYY-MM-DD)
• Keeps MEMORY.md under 200 lines as a clean index
• Preserves your writing style. Only the content changes, not the voice

AutoDream doesn't run after every conversation. It checks two gates: at least 24 hours since the last consolidation, and at least 5 new sessions since then. Both must pass. This means consolidation happens roughly once per day of active use: frequent enough to stay clean, rare enough to be invisible. When you would rather not wait, Consolidate now in the knowledge graph runs it on the spot.

• Time gate: minimum 24 hours between consolidations (configurable)
• Session gate: minimum 5 sessions since last run (configurable)
• Run it yourself with Consolidate now in the knowledge graph, which skips both gates
• Gate checks cost one stat + one directory scan. Negligible per-turn
• Never runs during sub-agent orchestration or background tasks

Every reply has a thumbs up and a thumbs down beside it. Thumbs down asks one question, what should have happened instead, and your answer is saved as a rule in project memory before you send the next message. Consolidation handles the slower half: when the same build breaks twice or you make the same correction twice, it writes the rule for you. That half arrives hours or days later, on the next run, not in the conversation you are having.

• Thumbs down with a line of explanation writes a rule immediately
• A bare thumbs is filed as evidence. A click alone does not say what the rule is
• Repeated failures need two occurrences before anything is written down
• Rules are markdown notes. Open the knowledge graph to reword the first line, or delete it
• 12 of the 40 index slots are reserved for rules, so newer notes cannot push one out

The consolidation agent is sandboxed. It can read and write memory files with plexon_read_memory and plexon_update_memory, and read the staged feedback signals with plexon_read_observations. No system commands, no code execution, no file operations outside the memory directory. A deadline ends the run, not a turn count.

• Three tools, all read-only or scoped to the memory directory
• No shell access, no file writes outside memory directory
• The observations tool exists only inside a consolidation run
• Timeout with graceful failure handling. A failed run retries sooner

The agent follows a fixed workflow: orient, consolidate, record relations, check its own evidence, prune the index, distil lessons, report. Each phase builds on the one before it. Phase 4 is the interesting one, because it makes the agent name the entry behind every change it wants to make and drop anything it cannot trace.

• Phase 1: Orient. List and read all memory files
• Phase 2: Consolidate. Merge duplicates, fix contradictions, update dates
• Phase 3: Relations. Cross-links, entity headers, supersedes
• Phase 4: Skeptic gate. Trace every write back to an existing entry
• Phase 5: Prune. Trim MEMORY.md, remove stale pointers, shorten verbose entries
• Phase 6: Lessons. Turn repeated failures and your feedback into rules
• Phase 7: Report. Brief summary of everything it changed

A PID-based lock file prevents multiple instances from consolidating simultaneously. If consolidation fails mid-run, the lock file mtime is rolled back so the gates retry sooner instead of waiting another 24 hours. The agent prefers editing over deleting. Information loss is treated as worse than verbosity.

• PID-based lock with 60-minute staleness timeout
• Mtime rollback on failure. No lost consolidation windows
• Prefers editing over deleting. Errs on keeping information
• Same provider/model as your main AI for prompt cache efficiency

Everything lives under Settings, Advanced, Memory. Turn consolidation off, or move the two thresholds until the cadence matches how you work. There is no manual trigger button. The gates are the only way a run starts, which is why they are the part you can adjust.

• Toggle on or off with immediate effect. No restart needed
• Configurable minimum hours and session count thresholds
• A separate weekly cadence for global memory, on by default
• Snackbar notification on completion: "Memory consolidated: N file(s) updated"