CASE STUDY · PUBLIC

memory-wire

The memory-bank CLI the operation writes durable cross-session lessons to: one bank per function, wired into the harnesses that need to remember.

Status ruling active development

Problem

Agents forget across sessions, and when they do remember, the memory is unstructured prose no tool can query. The operation needed memory as a first-class substrate with a CLI contract.

Build

  • Function-scoped memory banks: talent, growth, rd, ops, org - each written by the function's own routines, so cross-session lessons land where the next session looks.
  • A connect model that wires the banks into the agent hosts that exist on the machine, detected rather than configured.

Outcome

  • 607 of 608 tests pass at commit 34f03e6, summed across all nine test binaries. The one red is the Hermes golden-shape case: an environmental prerequisite on a machine without a Hermes install, expected green on an equipped host.
  • The 607-of-608 form is the honest one, and the ledger rule it taught is printed with it: the first pass summed two of nine binaries and read 584. An aggregate is a claim; re-derive it from the rows it names.

Provenance

Gate, executed607 of 608 (1 environmental red)cargo test --no-fail-fast at 34f03e6, 2026-10-21

Reproduce

Clone the repository, check out 34f03e6, and run `cargo test --no-fail-fast`, summing every binary's result line. On a machine with a Hermes install the golden-shape case is expected to pass; without one, the honest result is exactly 607 of 608.

From the documentation

  • A memory bank per function, so what an agent remembers is scoped by what that bank is for, rather than by a single undifferentiated store that grows until nothing in it is findable.

    BANK_IDENTITY.md, repository docs (not published: guard-excluded)

  • Retrieval efficacy is audited against the live banks as they fill, not against a benchmark, and the audit states that no ranking parameter, weight or threshold was changed by it, which is what makes its numbers measurements rather than tuning.

    RETRIEVAL_EFFICACY_AUDIT.md, repository docs (not published: guard-excluded)

  • The published suite figure carries its failed test with it (607 of 608) instead of being rounded to a pass, the same honesty rule the rest of the register follows.

    register row, 34f03e6