Skip to content
corpus.
How Corpus works

The system / Recall in practice

Recovering an investigation from a half-remembered question

A path through the records

From a fragment to an answer

An explanatory schematic of the edited transcript below. Each step links to its source section.

  1. A half-remembered questionThe Human has a question, but no record number.Read this step
  2. Search and reformulateBroad searches lead to more specific project names.Read this step
  3. Read sources and correctionsThe owning records carry qualifications worth recovering.Read this step
  4. A qualified, useful answerThe answer keeps the comparisons’ limits in view.Read this step

Recovering an investigation from a half-remembered question

An edited transcript walkthrough of an actual Corpus exchange, 18 October 2026. Not a replay, benchmark or recording of a fresh agent.

The question

The Human did not provide a record number:

reeves recall our investigations comparing t3 and omp and pi and that claw harness i forget

Reeve, the coordinating agent in this conversation, used the project’s record readers to recover the September investigations. The path was not a clean one-shot search.

Search, then reformulate

These are the actual search commands, in their original order:

python corpus.py recall t3 omp --limit 10
python corpus.py recall 'harness' 'pi' --limit 12
python corpus.py recall 'OpenClaw' --limit 10
python corpus.py recall 'oh-my-pi' --limit 8
python corpus.py recall 'pi-mono' --limit 8

The first response mixed unrelated mail and work records with the T3 investigation. The broad harness search surfaced general host infrastructure. Searching the project names more specifically found the OMP records and the Pi crossing. Search located candidates; it did not establish their conclusions.

Reeve then requested the owning records through story, including their corrections. Some returned output was truncated, and some records were read in the reader’s short form: this was useful retrieval, not a complete re-audit of every investigation.

Recover the corrections too

The source-reading sequence included:

python corpus.py story FIND-2798 FIND-2802 --full
python corpus.py story FIND-2649 FIND-2804 FIND-2827 --full
python corpus.py story FIND-2638 FIND-2639 FIND-2640 FIND-2648 FIND-2799 FIND-2800
python corpus.py story FIND-2641 FIND-2828 --full

Those corrections mattered. An earlier comparison had called message-delivery vocabulary missing from Corpus; the correction distinguished an existing design from its then-deferred implementation. Another corrected reports described as verbatim when they were condensed. A further correction narrowed an overly broad assurance: only the examined edits had been checked, not the entire report.

Recovering a conclusion without these qualifications would have recovered a more confident story than the record supported.

A useful answer, not a league table

Reeve identified the forgotten project as OpenClaw and recovered the September 23–24 comparisons. The inspected versions were T3 Code at 78af372c, Oh My Pi 18.2.5 at 37273117, Pi v0.87.1 at b4559750, and OpenClaw at 6ee9bc1b. These are historical inspection anchors, not claims about current releases.

The Human followed up:

did they find things we are better at

Reeve answered:

Yes—at the level of implemented structure, not measured overall performance.

The answer described the investigators’ favorable assessment of Corpus’s continuity across agent holders, evidence history, independent review and controlled integration. It also kept the crucial limit:

We had not demonstrated that those mechanisms produce better outcomes—or justify their cost—against the alternatives.

The learning went both ways. The recovered account named OMP’s practice of declaring what a shortened text discards, Pi’s framing rules exposing a Corpus JSONL defect, and OpenClaw’s shared wake throttling, loop limits and delegation caps as useful lessons. Identifying an idea to borrow is not evidence that Corpus subsequently implemented it.

What this example shows—and does not

This was an existing Reeve sitting with substantial conversation context. Reeve explicitly corrected the suggestion that it was a fresh, context-free successor. The historical investigation was retrieved from records, but this exchange does not isolate what prior context contributed.

No original screen recording was supplied or identified in the bounded source record. This page is an edited textual presentation of the exchange, with exact quotations and commands separated from editorial explanation. No new model runs were bought to recreate it; no latency or comparative performance result is claimed.

The useful result is modest and concrete: a half-remembered question led back to named investigations, their qualifications, and an answer the Human could use—without asking them to reconstruct the earlier work.