Synthesize Agent Memory Landscape

Synthesize the agent-memory landscape

Produce a public comparison whose numbers and qualitative findings come from one frozen population of main-analysis results.

Inputs and authority

Use the requested output, selected systems, and current or historical mode from the user request or invoking packet. A request to refresh a named artifact supplies its output authority. Without a file destination, return the synthesis in the response. Load the output collection's contract before writing there.

The evidence inputs are generated reviews under kb/agentic-systems/reviews/ and the exact retained results named by their analysis-result paths and analysis-result-sha256 values. Read the full retained results for findings: source register, shared records, memory/context lens, reconciliation, and limits. The compact review supplies publication identity and navigation. It cannot replace a missing full result or comparison assessment.

Use kb/types/agentic-system-analysis-result.md for the memory-comparison contract. Each row preserves its source revision, run, analysis cutoff, evidence tier, compared memory boundary, and per-axis assessment, basis, values, and canonical records. No legacy review, old CSV, transfer scan, or newly acquired source may supply or repair a finding. Missing required inputs block the selected population; report the main-analysis regeneration needed. Existing results must not be hand-patched to make a comparison pass.

Freeze the evidence

1. Select the population. Default a refresh to current inputs. Repeat --review to select the commissioned main reviews; omit it only when the commission covers all generated main reviews. Record the selection rule and exclusions. Select one review per source identity. A small selected set is a bounded comparison, with no implication of historical-corpus coverage.

2. Create a new bundle. Run from the repository root:

   uv run python scripts/bundle_agentic_landscape.py prepare --output <new-bundle-directory> --review <main-review-path>
   

   Repeat --review as needed. Add --ontology <kb/notes/path.md> for each additional ontology artifact actually used. The command reads main results directly, derives matrix.csv, and captures the exact review/result bytes, result contracts, reader code, producing and consuming instructions, and dependency declarations under their repository-relative paths. It writes snapshot.json and a canonical MANIFEST.tsv containing sorted sha256<TAB>byte-length<TAB>path rows. Require exit status zero and save the reported manifest hash outside the bundle before interpretation. Existing bundle directories are never replaced. The command does not update public matrix or table files.

3. Use only bundled evidence. Treat the bundle as immutable. If a needed finding or ontology input is absent, create a new complete bundle before drafting; do not mix in live files. For an existing historical bundle, verify it with its previously recorded manifest hash and use its bundled instruction and contracts. A method mismatch requires the matching pinned checkout. Legacy-corpus snapshots remain historical evidence, but this procedure does not rebuild or merge them into its population.

A temporary bundle suffices for a response or workshop trial. Before publishing, ensure the exact bundle and every cited original retained result are kept in Git, or identify a repository revision containing every input byte. A commit ID alone is insufficient when an input differs from that revision. Record the manifest hash, matrix hash, input file identities, and the reconstructable revision or retained-snapshot location in the published evidence boundary. A tracked comparison must remain auditable without ignored local run state.

Analyse and write

4. Compute quantitative candidates. Query the bundled CSV mechanically, decoding value cells as JSON arrays. For implementation/operation counts, use code-grounded rows with known values at wired, observed, or causally supported basis, plus absent assessments for evidenced negatives. Keep claimed and afforded findings separate. Keep doc-grounded findings in a separate qualitative section. Within each query, report inapplicable, uninspected, and not-determinable rows separately; none is an observed negative. A structurally valid unknown does not block unrelated findings.

   Retain an executable query and its output in a working query ledger. Each candidate names the fields, value-membership or set-equality test, tier and basis filters, numerator, denominator, included run IDs, and exclusions. Count each system once per query even when its value set contains several stores or routes. An assessed-subset proportion must name that subset; a whole-population prevalence claim requires complete applicable assessment. A change claim requires two verified snapshots, comparable scopes/contracts, and an explicit treatment of population changes.

5. Read and ground the mechanisms. For each selected finding, read the full bundled result and the cited canonical records, including their source evidence and limitations. Preserve the external mechanism and explain why the Commonplace term fits. Trace every qualitative example to a result path, hash, run ID, canonical IDs, and supporting section. Open-ended observations support named examples and contrasts, never prevalence from omitted mentions. Keep static wiring, observed use, contextual activation, and causal effect distinct. Withhold claims stronger than their records support.

6. Write one coherent snapshot. State the evidence identity, selection, source-tier population, source cutoffs, and analytical lens. Select only findings that the available population supports; do not pad a small pilot into a landscape survey. Give denominators beside numbers and scope beside comparisons. Link qualitative claims to their original retained result paths, using a section anchor where useful; compact reviews may additionally serve navigation. Do not cite the temporary bundle path. Name withheld conclusions and evidence gaps. Commonplace-specific recommendations belong in a separately commissioned transfer scan. Replace an incumbent synthesis as a complete snapshot, never by updating counts alone.

7. Verify the draft. Recompute every query from bundled bytes and check each example against its full result and records. If independent review is commissioned, give the checker the frozen bundle and expected hash, query ledger, and draft, without live corpus paths, transfer scans, or writer rationale. Otherwise perform these checks locally and report that mode.

8. Recheck and publish. Immediately before returning or writing, run:

   uv run python scripts/bundle_agentic_landscape.py verify <bundle-directory> --sha256 <recorded-manifest-hash> --source-root .
   

   For a historical snapshot, omit --source-root; always keep the externally recorded hash. The command checks the manifest, captured bytes, and exact matrix/result agreement. For current inputs it also checks source-file drift and population changes, including additions to an all-generated selection. On failure, withhold the draft and restart from selection. Write the commissioned output only after verification and evidence retention are satisfied; run commonplace-validate on every changed Markdown artifact. Public matrix/table refresh is a separate output: when commissioned, pass the identical explicit review list to both existing build scripts and check their recorded input identities against this bundle.

Report

Return the output path or response-only disposition; current or historical status; selection rule and source-tier population; cutoffs; manifest and matrix hashes; reconstructable evidence location; query verification and semantic verification mode; final bundle/source recheck; validation; and withheld claims. A fixture trial establishes procedure behavior, not external-system findings or production corpus coverage.