This is the operational status note for the OB1/OpenBrain to Honcho memory migration. The architecture decision is already covered in Agent Memory, Shared on Purpose. This page is the checklist: what we actually did, what proved out, and what changed when I called the migration operationally complete.
james-fleet-prod is the shared Honcho workspace for trusted persistent agents.38,813/38,813 queue rows processed, 0 unprocessed, 0 errors.deepseek/deepseek-v4-flash on Nous Portal/OpenRouter with reasoning disabled; summaries stayed on Fireworks DSF.38 curated records; 22/22 probes passed; July 8 smoke will email its report.This table is not a general leaderboard. It is the measured shape of this one migration workload: a large Honcho derivation queue with huge synthetic sessions, uneven summary prompts, and embeddings kept on OpenAI to avoid mixing vector spaces mid-import. The queue is now drained; the table preserves the routing decisions that got it there.
| Route | List price used | Observed migration behavior | Final call |
|---|---|---|---|
| Nous Portal / OpenRouter DSF | deepseek/deepseek-v4-flash$0.09/M input, $0.18/M output | Retained deriver route. Clean final proof: 38,813/38,813 rows processed, 0 unprocessed, 0 errors. Reasoning was explicitly disabled. | Winner for this backfill: cheap enough, clean enough, and operationally simple. |
| Fireworks DSF | deepseek-v4-flash$0.14/M input, $0.28/M output | Safe fallback and retained summary route, but later deriver samples were much slower than the Nous DSF tail. | Keep as fallback / summary lane, not the fastest deriver. |
| OpenAI API | gpt-5.4-mini$0.75/M input, $4.50/M output | Fastest option in earlier samples, but expensive for this backlog and less attractive once cheaper DSF drained cleanly. | Speed candidate only; not the cost-control route. |
| Free / subscription candidates | StepFun free, Tencent Hy3 free, Gemini Flash Lite, Grok/SuperGrok OAuth | All were canaried. StepFun caused repair/zero-observation issues; Hy3 stalled in live drain; Gemini worked but was not fast enough; Grok required a fragile proxy path and did not beat DSF. | Rejected despite attractive headline pricing. |
Final proof basis: live Honcho queue snapshot on July 7, 2026 after the Milo/OpenClaw bulk backfill drained: 38,813 total rows, 38,813 processed, 0 unprocessed, 0 errors. Embedding costs are still excluded; the point of this table is operational routing, not a universal model leaderboard.
The important lesson after the queue drained: bulk derivation is not the same thing as usable continuity. The repair pass converted the highest-value Milo/OpenClaw history into direct, answer-shaped memory records and then tested normal recall, not just database counts.
| Pass | What changed | Recall result |
|---|---|---|
| Core continuity | 23 records for identity, family, relationships, infra pointers, and canonical facts: Alyx, Cindy proposal, Nancy/Dory, Aurora, Vaultwarden pointer, Forge, DGX Sparks, MiloBridge, and Karpathy Loop. | 10/10 probes passed after targeted repairs. |
| Project/wiki continuity | 15 records for Angels vs Demons, Milo-Ark, StackChan, PuppyPi, Echo, Bandit, Family AI Fleet, OpenClaw Lab Node, and the local-first voice pipeline. | 12/12 probes passed after one voice-pipeline repair. |
| Secret handling | Both bundles used pointer-only credential handling and explicit redaction policy. | 0 secret-scan hits before import. |
| Human-readable audit | A curated Obsidian note was written to the AI Inbox so the migration has a readable proof record outside the agent memory system. | Done; delayed recall smoke scheduled for July 8 and routed to email the report. |
This is deliberately small. The point is not to stuff Honcho with every old log; it is to preserve the facts future Milo sessions should answer correctly without dragging raw archive noise into every turn.
| Step | Status | Result |
|---|---|---|
| Export and classify OB1 memory | Done | Operational and personal rows were separated from blocked, review, raw, stale, and superseded material. |
| Build production allowlist | Done | The shared import used the approved bundle only: 567 operational rows and 58 personal rows. |
| Create shared Honcho workspace | Done | james-fleet-prod became the single production memory target for trusted persistent agents. |
| Import approved data | Done | Approved data landed in Honcho; blocked/review/raw lanes stayed out. |
| Drain OB1 derivation queue | Done | Honcho processed the approved OB1 import and later canary/watchdog writes back to zero backlog. |
| Drain Milo/OpenClaw history backfill | Done | The July 3 comprehensive backfill finished on July 7 with 38,813/38,813 queue rows processed and 0 error rows. |
| Distill continuity seed records | Done | Two answer-shaped seed passes imported 38 curated records into james-fleet-prod instead of relying on raw transcript recall. |
| Run recall probes | Done | Core and project/wiki probes passed 22/22 after small targeted repairs; a delayed July 8 smoke will check stability, write a markdown report, and email it to James. |
| Cancel provider-compartment PR | Done | PR 54534 proved the seam, but the live design intentionally became simpler: one shared workspace plus write hygiene. |
| Cut over default Hermes/Miloh | Done | The default Hermes path writes to james-fleet-prod and produced a searchable cutover marker. |
| Cut over OpenClaw/Milo | Done | Milo kept its existing OpenBrain-style tools while a local compatibility shim routes the backend to Honcho. |
| Start shared watchdog | Done | The old ops-only watchdog was removed; the remaining shared watchdog runs every four hours and writes directly to Honcho. |
| Retire split canaries | Done | honchoops, honchopersonal, and the temporary honchofleet profile are archived; OpenClaw keeps only the live Honcho shim. |
| Update public notes | Done | The blog now says the migration is complete for live trusted-agent paths; compartments are historical proof, not the active plan. |
| Unwire OB1 MCP from live Hermes | Done | Default Hermes and non-archived canary profiles no longer expose ob1-memory; stale prompts telling agents to call ob1_writeback were removed. |
| Proof | Evidence | Meaning |
|---|---|---|
| Default-path write/search | Hermes and OpenClaw cutover markers are searchable in james-fleet-prod. | The two live paths are not merely configured; they wrote through to the new backend. |
| Scheduled watchdogs | Direct watchdog marker HONCHO_DIRECT_WATCHDOG_20260702T171800 wrote/search-verified through Honcho, and the forced cron run passed. | The backend path is now tested without depending on a throwaway Hermes canary gateway. |
| Queue recovery | July 2 cleanup state after direct watchdog verification: 72/72 completed, pending=0, in_progress=0. July 7 live state after the Milo/OpenClaw backfill: 38,813 total queue rows, 38,813 processed, 0 unprocessed, 0 errors. | Both the OB1 cutover and the later history backfill drained cleanly. |
| Recall continuity | Imported 23 core continuity records and 15 project/wiki records, then ran normal Honcho reasoning probes. Final pass: 22/22 probes passed after targeted repairs. | The migration now has usable answer-shaped continuity, not just a completed bulk queue. |
| Public audit trail | Curated audit note written to Obsidian AI Inbox and local work logs updated with bundle paths, queue checks, and repair notes. | James and future agents have a readable proof record outside Honcho itself. |
| Archive-only preservation | OB1 material remains historical provenance, but the ob1-memory MCP server is no longer wired into normal Hermes sessions. | The safety record exists without running a second active memory system day to day. |
| Remaining item | Status | What would close it |
|---|---|---|
| Recall-quality tuning | Optional follow-up | Keep watching real conversations for noisy retrieval. Early family-memory checks recovered Nancy context but still showed thin/uncertain Alyx context, so promote only high-confidence facts. |
| OB1 archive hardening | Optional follow-up | Rotate/revoke old OB1 credentials and keep any exports/backups as offline provenance once no emergency rollback path is needed. |
| Delayed recall smoke | Scheduled July 8 | Re-run canonical probes after Honcho settles. The cron job writes the report locally and emails it to James; local cron history remains backup, not the notification path. |
| Public/current-state docs | Done | This page, the Obsidian audit note, the local work log, and the migration skills now agree: shared Honcho is live; compartments are historical proof; continuity recall has been seeded and tested. |
0 unprocessed rows.james-fleet-prod.Done: import, shared workspace, default Hermes cutover, OpenClaw/Milo shim, direct watchdog, split-canary cleanup, active OB1 MCP unwiring, Milo/OpenClaw bulk backfill, continuity seed repair Current state: Honcho healthy, shared memory live; bulk queue drained to 38,813/38,813 processed with 0 errors; 38 answer-shaped continuity records imported; 22/22 recall probes passed after repairs Still deferred: optional OB1 credential/archive hardening; delayed July 8 recall smoke will email its report; future repair only for real misses Decision: migration operationally complete; no rollback path is being kept warm; do not broad-import raw logs just because they exist