AkitaOnRailsandClaude Fable 5
0b35ed873e
Every user-facing doc was graded on four axes (what / why / when-to / real example, including when-NOT-to). Fixes for the highest-impact gaps: - managed-workstreams: 'Do you need this?' - hooks + handoffs already cover the quit-Claude-open-Codex case; managed runs are for native resume surviving a harness switch. Skip guidance included. - auto-improvement-loop: a 60-second user-facing top for a feature that is on by default - what auto-approve means, the explicit require_approval=true recommendation for shared/team servers, cost shape, and what a staged proposal sidecar contains. Research prose retained below the fold. - okf: opens with the user payoff (hand a bundle to someone without ai-memory; export-okf one-liner; what the receiving side sees) before the conformance design. - temporal: when to reach for as_of (audits/post-mortems; plain memory_query is right 99% of the time) plus a worked postgres-then-migrated example with both results. - typed-edges: the gotcha->fixes->contradicts->lint loop as a concrete scenario, and 'plain wikilinks remain the default' skip guidance. - experience: a sample staged cross-session proposal. - config template: consolidation and slots blocks say why/when, not just what. Accuracy bugs found by the same sweep: auto-scope's config snippet still called mode="single" the default (per_actor since v1.39); ZCode was listed twice with conflicting statuses in both support matrices (it has MCP #529 AND hooks #532 - merged to one Supported row); the README docs index listed ROADMAP-2.0 twice. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MDbhmszrjG9s5MrPrTuNtm
3.0 KiB
The experience pass (cross-session abstraction)
2.0 item 6. Per-session auto-improvement reviews one trajectory at a time — it can never see that four sessions repeated the same workflow, or that the operator keeps re-stating a preference no page records, or that recent sessions quietly contradict a stored decision. The experience pass reads the last N session summaries of a project side by side and proposes exactly that cross-trajectory knowledge.
Why "experience"
It is the Reflection→Experience step the 2026 agent-memory literature converges on (the survey's third stage; TriMem's narrative layer): raw capture → per-session summaries → durable knowledge distilled across trajectories. The pass targets pattern/procedure/preference/ architecture pages — the pages that make session #50 cheaper than session #5.
When to turn it on — and when not to
Opt-in, off by default:
[auto_improve.scheduler]
experience_every_sessions = 5 # run after every 5 newly completed sessions
experience_sessions = 10 # read the last 10 session summaries
Turn it on when per-session auto-improve is already working for you and the project has real session history (the pass skips scopes with fewer session pages than the cadence floor). Leave it off when the project is young — cross-session patterns need sessions to cross — or when no LLM is configured: the pass is LLM-hosted and the zero-LLM default path never runs it.
What a proposal looks like
After eight sessions in which the operator repeatedly rebuilt the release binary before running the eval harness, the pass might stage:
# _pending/auto-improve/…-procedures-eval-run.md (confidence 0.86)
## Proposed: procedures/eval-run.md
Before benchmarking, rebuild the release server first — the harness
spawns `target/release/ai-memory`, and a stale binary silently
benchmarks old code.
cargo build --release -p ai-memory-cli
cargo run --release -p ai-memory-eval -- retrieval
Evidence: sessions 2026-08-28 ("rebuilt release, numbers changed"),
2026-08-30 ("forgot the rebuild again — rerun").
It stays in _pending/ (with require_approval = true) until a human
approves it — the same review flow as every other auto-improve
proposal.
What it costs and what guards it
One LLM call per project per cadence trigger, prompt-bounded like the
per-session reviewer. Every proposal flows through the identical
machinery: JSON-schema constrained output, validation, the confidence
floor, the eval gate, the rejection buffer (rejected ideas are shown to
later runs so they are not re-proposed), and pending-writes staging with
sidecars — reviewable, never silent. require_approval applies
unchanged. The system prompt demands evidence spanning at least two
sessions, naming them; single-session findings are the per-session
reviewer's job and are rejected here.
The cadence anchors on enablement: switching the pass on against an old
store does not re-digest history — it waits for the next
experience_every_sessions completed sessions.