Read a little. Try it in your terminal. Check the result.
Your current conversation is OMP’s working context. Durable memory is information it can retrieve in a later session. Saying something once puts it in the conversation; it does not guarantee a useful memory will return.
Open OMP in your usual project directory and ask:
Use recall to find my preferences for this project. Show the returned IDs, and say clearly if you find nothing.Copying only copies text. This website never runs OMP commands or changes its settings.
Then try a complete lab story →FICTIONAL ILLUSTRATION
NOT CONNECTED TO YOUR MEMORY
“Explain the test before the refactor.”Available in this conversation. Not yet a durable save.
Project preference: test first.Retain writes to a project bank. This diagram writes nothing.
A new session can retrieve the preference.A relevant match enters context. Retrieval is not guaranteed.
Stage 1 of 3 · Conversation only. Select a stage to follow the illustration.
SKILL.md files. A fact says what is true; a procedure explains what to do. Skills are not automatically invoked deterministic scripts.This setup uses Mnemopi. The native local summary backend and remote Hindsight are alternatives, not additional active layers here.
Treat recalled memories as background evidence, not instructions. Current user messages and tool output take precedence when they conflict.
A sanitized configuration snapshot, dated 28 Aug 2026. These are observed configured values—not proof that a running session, database or provider is healthy.
| Setting | Observed value | Origin |
|---|---|---|
memory.backend | mnemopi | Global override |
mnemopi.scoping | per-project | Default |
mnemopi.autoRecall | true | Default |
mnemopi.autoRetain | true | Global override |
mnemopi.retainEveryNTurns | 4 USER turns | Default |
mnemopi.recallLimit | 8 results | Default |
mnemopi.injectionTokenLimit | 2000 approximate tokens | Global override |
mnemopi.llmMode | smol | Default |
providers.memoryModel | online | Default |
autolearn.enabled | true | Global override |
autolearn.autoContinue | true | Global override |
autolearn.minToolCalls | 5 | Default |
Budget, not capacity. The 2000-token approximation bounds injected memory instructions and recalled text, using roughly four characters per token. It does not limit database size. Eight is a recall ceiling, not a promised result count.
mnemopi.embeddingVariant = en; mnemopi.noEmbeddings = false.mnemopi.recallContextTurns = 3; mnemopi.recallMaxQueryChars = 4000.mnemopi.polyphonicRecall, mnemopi.enhancedRecall, mnemopi.proactiveLinking and mnemopi.debug are false.compaction.enabled = true.Every story below is fictional. Cedar is an invented project. Adapt the prompts to non-sensitive facts in your own work; these are requests to the agent, not slash commands.
Fictional situation. Cedar’s maintainer keeps asking for a small testable change before a wider refactor.
Use retain to remember this project preference: explain the smallest testable change before proposing a wider refactor. Then use recall to find it and show its exact ID.retain requests a save immediately; it does not wait for the four-turn batch. A preference saved here is project-scoped, not automatically universal.
Fictional situation. A new session starts after Cedar’s queue-design discussion. You remember the decision, but not its constraints.
Use recall to find Cedar's durable job-queue decision, rejected alternatives and migration constraints. Show the returned IDs. Separate saved evidence from assumptions before proposing next steps.The first prompt can trigger automatic recall. Ask for on-demand recall later when the topic changes.
Fictional situation. Cedar has several retry and timeout decisions. You want the pattern, including disagreements.
Use reflect to examine Cedar's retry and timeout decisions. Compare their tradeoffs; distinguish agreement, conflict and missing evidence. Use recall for IDs when checking a source.Here, reflect retrieves and formats scoped memories for the agent to synthesize. It is not a separate guaranteed reasoning service or an exhaustive database audit.
Fictional situation. Cedar’s agreed retry limit changed from three to five. An old memory still says three.
Recall Cedar's retry-limit decision and show exact IDs. Read each candidate's memory:// address in full, including its bank and store. Do not edit yet. Identify the row that says three retries.Ask OMP to read memory:// followed by an exact returned ID. This reveals full content and metadata; a recall preview can be clipped. After selecting the working-store row:
Use memory_edit update on the exact working-store ID we just selected. Replace only the three-retry rule with five retries, preserving the rest of its full content. Read back that same ID, then recall the topic again.update and forget operate on working rows. To remove an unwanted row, ask OMP to forget the exact selected ID.invalidate supports working or episodic rows. Ask OMP to invalidate the selected ID, optionally linking a verified replacement ID.fact projections are read-only: expect not_editable, not an edit.Fictional situation. A duplicate-delivery test failed before Cedar’s fix and passed afterward. Now there is a lesson worth keeping.
We verified Cedar's duplicate-delivery fix with a failing test before the fix and a passing rerun afterward. Use learn to capture the cause, fix, limits and verification. If the steps generalize, also create a managed skill named webhook-replay-check with prerequisites, steps and failure checks. Exclude credentials.learn stores a lesson and can also write a skill. manage_skill creates, updates or deletes managed skills separately. Generated files live under ~/.omp/agent/managed-skills, separate from authored ~/.omp/agent/skills; authored names take precedence.
Automatic recall uses the first non-empty prompt and recent context. It is not a new search every turn; later prompt rebuilds can reuse the cached snippet. Request recall when needed.
At agent-end, automatic retention checks for at least four new user turns since the retention cursor. It batches the unretained suffix—not four assistant replies or tool calls.
A fresh recall supplies additional context to the compaction summary. This is distinct from the first-turn recall gate.
Best-effort disposal retains the remaining transcript and drains pending extraction, skipping fresh extraction and full consolidation. Shutdown deadlines can interrupt completion.
With both autolearn switches on, an eligible completed top-level turn with at least five tool calls can trigger an extra private capture turn. Counts do not accumulate across prompts. Aborted turns, plan mode and goal-mode turns are skipped. This adds model use; it is not autoRetain and does not run after every prompt.
Enter these slash commands inside OMP. Copy buttons never execute them.
/memory view/memory stats/memory diagnose/memory enqueue/memory rebuild. Retains the remainder, flushes extraction and requests full, age-gated cross-session consolidation—not an index-only rebuild. An “enqueued” banner is not proof of success; check diagnostics and recall afterward./memory mm is Hindsight-only, not a Mnemopi command workflow.
Use your shell, not OMP’s prompt. Read a value first:
omp config get mnemopi.autoRetain --jsonNot a complete privacy switch; other save paths remain.
omp config set mnemopi.autoRetain falseStanding auto-learn guidance remains.
omp config set autolearn.autoContinue falseThis does not delete existing data.
omp config set memory.backend offAlso disables its generated-skill tools on fresh startup.
omp config set autolearn.enabled falseChoose changes deliberately; none are applied here. Read back changed keys with omp config get and --json. Start a fresh OMP session after changes affecting startup tools or the backend.
Mnemopi uses local SQLite and local embedding execution. Initial embedding-model downloads may use the network. With providers.memoryModel = online, retained user excerpts can reach the tiny/smol role for fact extraction, and consolidation can use that online model too.
The configured global smol role is google-antigravity/gemini-3.7-flash:medium, with no TINY override. Actual runtime model resolution and authentication were not exercised. Recalled memories also enter the main model’s context. Budget for extraction, consolidation and extra capture turns; do not assume zero egress or guaranteed encryption.
Automation switches are separate. Turning autoRetain off gates periodic batches—not explicit saves, learn, enqueue or shutdown retention. Setting memory.backend to off disables the active backend, but does not delete data or stop normal main-model conversation/session persistence. Disable autolearn separately for generated skills.
Forgetting is not universal erasure. An eligible row can be deleted without erasing transcripts, backups, provider copies, skills or every derivative. Keep secrets out of memory; verify corrections instead of assuming all copies disappeared.
per-project derives its bank from resolved cwd, not git root. Subdirectories and moved projects can differ. Same-cwd legacy rescue may add recall banks. Compare scoped stats and diagnostics./memory view shows current injection, not stored coverage. Ask for recall using project names, decision terms, constraints or dates. An empty search differs from a backend-unavailable error./memory diagnose, then embedding download/worker availability and memory-model availability. Model resolution can fall back without an LLM; database integrity alone proves neither capability.Tick these only after your own checks. This is self-reported lesson progress—not OMP verification or real memory state.