Skip to content

CHANGELOG.md

Latest commit

 

History

History
259 lines (209 loc) · 12.4 KB

CHANGELOG.md

File metadata and controls

259 lines (209 loc) · 12.4 KB
Raw

Changelog

All notable public-facing changes to AIppocampus are summarized here. The project is still alpha; release notes distinguish shipped behavior from evidence, diagnostics, and claim boundaries.

Unreleased

The next slice is about fewer doors that open onto machinery.

  • First recall now has one visible next move. Provider status, health/status, MCP tools, latest-reply, and compact recall prefer a small action card over a wall of flags or handles.
  • Setup paths are kinder. Plugin install, local sync, object-sync, provider-key, logs, doctor, storage GC, and conversation import explain status, plan, apply, and no-write boundaries before they ask for paths or credentials.
  • Navigation-only features say so out loud. Warm ambient recall, self-notes, action-time hints, Telepathy, question tracking, and the cognitive-map frontdoor can guide attention without pretending to be source truth.
  • Release and benchmark doors are calmer. Focused tests run first, pr is the local pre-push gate, benchmark smoke stays claim-boundary aware, and compact summary JSON exists for agents that only need the verdict.
  • The docs now lead with the first useful thread coming back into view, not a readiness ceremony.

Still alpha. Still source-backed. A little less control room, a little more campfire map.

0.3.3 - 2026-06-15

0.3.3 is a dogfood release for fewer sharp edges.

AIppocampus should feel easier to keep nearby: quieter status, clearer routes, less release-theater, and fewer places where an agent has to guess whether a diagnostic is evidence.

  • Recall gets better handholds: source-backed aliases, safer candidate planning, outcome feedback, route labels that survive packet trimming, and an attention-router guard that will not promote a less relevant route just because it scored loudly.
  • MCP and public outputs are more careful by default. Search is metadata-only unless snippets are requested, thread ids are bounded, storage dry-runs are public-safe, and sync status gives copyable aippocampus commands.
  • Install/update is less fussy. apply --all-local can refresh the local Codex plugin cache when there is one clear target, and status can acknowledge when the foreground agent already sees the tools.
  • The test and release path is calmer: focused tests first, pr at most once locally, and release preflight for the few checks that can drift after CI.
  • Benchmark and doctor reports are smaller and more honest. A report existing is no longer confused with a public-quality claim.

Still alpha. Still source-backed. Fewer knobs in the doorway.

Compare

0.3.2 - 2026-06-15

0.3.2 is a cleanup release for trust at the edges.

The main change is not louder memory. It is calmer proof. Install, export, maintenance, benchmark, and release paths now say more clearly what happened, what is safe to share, and what still needs a human or operator look.

What Feels Different

  • Release checks are less ceremonial. The planner now points to the checks that matter for the changed surface, and local preflight no longer tries to be CI.
  • Codex plugin install success is readable at a glance. --json now returns a compact success summary; --operator-json keeps the full deep-debug report. Rollback also has a dry-run preview.
  • Public issue attachments have a real metadata-only public-export path. Clean-source text, session ids, anchors, graph labels, and searchable SQLite indexes stay out.
  • Search, recall, MCP tool listing, health, and CLI help now have quieter public-safe modes for agents that need the next step, not the whole machine.
  • Expected operator failures return structured JSON and useful exit codes instead of tracebacks.
  • Maintenance and storage reports are bounded by default, with full audits kept explicit.
  • Warm ambient status now separates queued work from actual worker evidence.
  • Benchmark reports are more honest about proxies, fixture gates, and public quality claims.

Still alpha. Still source-backed. Less smoke, clearer lantern.

Compare

0.3.1 - 2026-06-15

0.3.1 is a small release about foreground manners.

AIppocampus can now let architecture-native route work be felt without turning the frontstage into a wiring diagram. When attention-router, macro, topology, or local/global cues matter, agents get a tiny navigation hint and the same old boundary: deepen before claims.

Human recall output also stops printing long opaque navigation handles by default. The handle is still there for JSON and MCP callers; people get a short next action instead.

Still alpha. Still source-backed. Less transport noise, more useful handhold.

Compare

0.3.0 - 2026-06-15

0.3.0 is the release where AIppocampus becomes easier to invite into an agent's day.

The promise is the same: continuity should come from source trails, not from a model pretending it remembers. What changed is the feel. The package now gives agents clearer doors to knock on, clearer reasons to stay quiet, and a much shorter path from "is this installed?" to "yes, here is the next source-backed step."

What Feels Different

  • First run is calmer. The Codex plugin install path, update status, onboard status, sync status, and nested CLI help now speak in one public language.
  • Agents get better handholds: recall, AIppo, deepen, and explain are exposed as agent-native affordances instead of hidden machinery.
  • Prompt-hook recall is more polite. Small casual prompts stay cheap; explicit continuity prompts route toward the right next action.
  • Successful plugin verification is now readable at a glance with --compact-json, without local paths or unrelated host noise.
  • Benchmark and evidence work is better organized, with stronger gates against turning diagnostics into grand claims.

Still Alpha

  • 0.3.0 does not change the package classifier. AIppocampus remains Development Status :: 3 - Alpha.
  • It does not claim universal recall quality, hosted-service maturity, all-client coverage, or a stable Python SDK.
  • Live pilots, benchmark reports, Dream/macro/Telepathy work, and LongMemEval source-side evidence are useful signposts. They are not a victory lap.

Compare

0.2.0 - 2026-06-07

0.2.0 is the first broad continuity release after the 0.1.x metadata and discovery repair tags. Before this release-prep change, the mainline range from v0.1.1 through PR #889 covered 392 commits touching 802 files, with 252 merged PRs after the v0.1.1 release PR.

Highlights

  • Reframed AIppocampus around source-backed continuity, not generic task memory. The installable skill, root agent guidance, and public-facing docs now lead with clean source, source refs, packet trust, and the difference between navigation scent and evidence (#805, #806, #807, #808, #812, #816).
  • Added recall packet trust/action grammar: direction_only, reopenable_route, bounded_evidence, source_open, and ignore_or_blocked. Foreground packets can now stay precise about what an agent may do next without pretending every cue is proof (#761, #770, #788, #798, #839).
  • Advanced fresh-thread and ambient recall from demo fixtures toward a more usable runtime path. This includes fresh-thread action contracts, living cue cache selection, prompt-hook consumption, hot-path recall intent guards, active path packets, magic-preserving warm activation, and the #281 field continuity readout (#497, #508, #606, #632, #654, #731, #770, #814, #889).
  • Expanded the source-backed evidence surface with new or hardened benchmark families: multimodal corpus/media/NIAH fixtures, Field Continuity, hippocampal recall and hard negatives, segmented merge policy, E2E50 silent constraints, semantic robustness Track S, AMemGym, MemoryAgentBench, and benchmark priority/claim-boundary maps (#494, #504, #529, #537, #538, #539, #553, #562, #565, #572, #730, #737, #738, #748, #774, #885).
  • Added stronger recall navigation, route diagnostics, and repo familiarity paths. The runtime now has source-joined routing diagnostics, navigation potentials, source texture sidecars, repo familiarity affordances, issue-route quality smoke, and live semantic route actionability readouts (#551, #552, #628, #655, #767, #813, #845, #855, #875, #876, #878).
  • Added coding-continuity read models and integrity guards, including SequencePacket/Episode/Arc evidence, code-state anchors, correction host-event capture, weak-covered operation validation, conflicting operation diagnostics, and an operation claim integrity gate (#766, #779, #857, #861, #862, #874, #886, #887).
  • Deepened Dream/Journey and subconscious work from planning notes into bounded runtime and evidence slices: Dream manual review ingestion, user-visible eval axes, trust horizon gates, Journey resonance, working-memory compaction, delivery prefiltering, source-body stance/atmosphere arcs, constructive invitations, bridge hypotheses, and coding Dream probe selection (#613, #618, #619, #629, #645, #648, #717, #844, #871, #880, #881, #888).
  • Hardened storage, sync, and maintenance internals. This includes storage GC, generation pointer publishing, reader pins, encrypted sync v2 boundaries, migration recovery diagnostics, key-provider contracts, encrypted sync head graph conflicts, provider metadata evidence, local update status/apply, and preemptive lifecycle health/snapshot paths (#524, #636, #686, #705, #708, #719, #720, #721, #722, #723, #724, #728, #729, #740, #741, #778, #841, #842).
  • Improved public package, CI, and release hygiene. Python support is now explicitly 3.12/3.13, deterministic benchmark smoke and optional OpenAI Agents SDK smoke run in CI, maintainer shipping lanes and release checklist are documented, dependency and safe-environment contracts are clearer, MCP registry readiness checks were repaired, public client matrix evidence was refreshed, and CodeQL alert surfaces were cleared (#474, #495, #496, #499, #509, #510, #511, #687, #692, #702, #883).
  • Removed resolved package-only and flat compatibility shims after the package entrypoints and public CLI facade were made clearer. Use the packaged aippocampus CLI and module-owner entrypoints rather than old flat helper scripts (#479, #480, #547, #548, #549, #550, #596, #677, #685).

Compatibility Notes

  • Public package metadata now declares version 0.2.0.
  • Python 3.12 remains the support floor; CI/package metadata cover Python 3.12 and 3.13. Python 3.10 and 3.11 are not public support targets.
  • The runtime still has no mandatory third-party runtime dependencies. Release, development, benchmark, and optional OpenAI Agents extras remain opt-in.
  • Legacy helper shims have continued to shrink. Scripts and docs should prefer aippocampus ... or package-owned modules over old flat helper paths.

Known Boundaries

  • This changelog prepares the 0.2.0 release notes; it does not create the tag, publish PyPI, update MCP Registry metadata, or create a GitHub release.
  • At the time of the 0.2.0 preparation notes, the release checklist still expected a broad local evidence sweep. Current releases should use python tools/aippocampus/test_plan.py --release-preflight --json and docs/guides/setup/release-checklist.md instead of replaying the historical quick/pr/benchmark/coverage/full stack by reflex.
  • Current evidence does not claim interactive Codex Desktop marketplace click-through, third-party install review, macOS/Linux signed binaries, broad provider/client coverage, or production-quality private-history recall.
  • Many benchmark and smoke surfaces are intentionally contract smokes or diagnostic proxies. They should not be summarized as live model quality, private-history quality, or broad superiority claims unless a dedicated dated evidence owner says so.

Compare

0.1.1 - 2026-06-02

  • Metadata repair release tag for the package and discovery surface.

0.1.0 - 2026-06-02

  • Initial public package/discovery tag.