v0.36.0 - A1 server mode

A1 Server Mode - hippo serve, MCP-over-HTTP, thin-client CLI

Added

• hippo serve daemon on http://127.0.0.1:6789 (configurable via --port or HIPPO_PORT). Exposes /v1/memories, /v1/auth/keys, /v1/audit, MCP-over-HTTP at /mcp, and /health.
• CLI thin-client. When the daemon is running, CLI invocations auto-detect via .hippo/server.pid and route through HTTP. Stale pidfile self-heals on first ECONNREFUSED.
• MCP-over-HTTP/SSE transport alongside the existing stdio path. POST /mcp synchronous JSON-RPC; GET /mcp/stream SSE keepalive.
• Domain layer src/api.ts. Pure functions for remember/recall/forget/promote/supersede/archiveRaw/auth*/audit. Both server and CLI delegate through this surface.
• HTTP auth middleware. Bearer token via Authorization header; loopback (127.0.0.1, ::1, ::ffff:127.0.0.1) accepts unauthenticated as actor='localhost:cli'. Non-loopback no-token returns 401. Server refuses to bind 0.0.0.0 without auth.
• 24h soak harness at benchmarks/a1/soak.ts (manual run).
• p99 recall benchmark at benchmarks/a1/p99-recall.ts.

Fixed

• Audit-log tenant attribution: audit() helper now uses entry's tenant_id instead of HIPPO_TENANT env (latent bug).
• api.archiveRaw and api.forget enforce tenant scope: cross-tenant returns "memory not found".
• SIGTERM drain: server.closeAllConnections() before close so SSE streams don't block shutdown.
• MCP-over-HTTP threads hippoRoot + tenantId from auth context.

Internal

• 99 new tests (730 -> 829 + 2 skipped). Headline parity test spawns real subprocess server; concurrent recall+write under SQLite single-writer (10 readers + 1 writer) confirms zero locked errors.
• All 5 /review ship blockers closed before merge.

Known issues (tracked for v0.37.0 in TODOS.md)

• p99 latency: 58.4ms vs 50ms target on 10k store. Architecture ships; latency hardening lands in v0.37.0.
• HIPPO_API_KEY silently dropped on stale-pidfile fallback.
• Concurrent hippo serve has no winner detection.
• Recall ?mode=hybrid accepted but ignored (BM25-only over HTTP).
• MCP-over-HTTP SSE is keepalive-only; no server-pushed messages.

🤖 Generated with Claude Code

v0.35.0 - A5 stub auth

A5 Stub Auth - API keys, audit log, tenant isolation

Added

• Schema v16: tenant_id on memories, working_memory, consolidation_runs, task_snapshots, memory_conflicts (default 'default') + composite indexes. New tables: api_keys (scrypt-hashed), audit_log (append-only).
• API key primitives: createApiKey / validateApiKey / revokeApiKey / listApiKeys. scrypt + timingSafeEqual. Plaintext returned exactly once.
• Audit log: hooks on remember, recall, promote, supersede, forget, archive_raw, auth_revoke. Query via hippo audit list --op X --since Y --json.
• Tenant resolution: resolveTenantId({db?, apiKey?}). Order: api key > HIPPO_TENANT env > 'default'.
• Cross-tenant isolation: enforced on CLI recall/explain/context, MCP server (hippo_recall, hippo_context, hippo_status), dashboard.
• CLI: hippo auth create [--label X] [--tenant Y], hippo auth list [--all], hippo auth revoke <key_id>, hippo audit list [--op X] [--since Y] [--limit N] [--json].
• SSO/SCIM stubs: throw NotImplementedError, tracked for v2.

Fixed

• Empty HIPPO_TENANT env trims to 'default'
• bigint-safe JSON for audit metadata
• archiveRawMemory audit event uses row's tenant_id, not env

Internal

• 36 new tests (730 -> 766). Cross-tenant negative test verified at vitest layer + live dashboard API smoke test.
• All /review findings closed: 4 HIGH (tenant filter holes on MCP/explain/dashboard/context), 7 MEDIUM, 8 LOW.

Deferred to v2 (TODOS.md)

Multi-tenant per-key isolation, OAuth/SCIM, audit log retention, RBAC, background pipeline tenant scoping (consolidate, embeddings, autolearn, autoShare).

Generated with Claude Code

v0.34.0 — A3 provenance envelope + pineal salience v2

0.34.0 (2026-04-29)

Added

• A3 provenance envelope. Every memory now carries kind (raw | distilled | superseded | archived), scope, owner, and artifact_ref columns. hippo recall --why surfaces the envelope; hippo remember accepts --kind, --scope, --owner, --artifact-ref flags. See MEMORY_ENVELOPE.md.
• Append-only invariant on kind='raw'. SQLite trigger trg_memories_raw_append_only aborts direct DELETE on raw rows. The only legitimate path is archiveRawMemory(db, id, { reason, who }) which snapshots into the new raw_archive table, purges the FTS row, and removes the memory in one SAVEPOINT (sets up A4 right-to-be-forgotten).
• Schema v14 + v15. v14 adds the envelope columns, the raw_archive table, the append-only trigger, and INSERT/UPDATE CHECK-substitute triggers (ALTER TABLE cannot add CHECK in SQLite). v15 closes a NULL-kind bypass in those triggers and adds UNIQUE(memory_id, archived_at) to raw_archive. Backwards compatible, auto-migrates.
• Pineal salience v2. --salience-threshold flag for the recall pipeline (commit 50528a5).
• Enterprise execution roadmap (ROADMAP-RESEARCH.md). 90-day plan re-sequenced after Codex + eng-review pass: A3 envelope first (this release), then A5 stub auth, A1 server, E1.3 Slack ingestion. Cuts 7 deferred items into days 91-180.

Fixed

• FTS leak in archiveRawMemory. Archived raw content stayed in memories_fts until next DB-open backfill; defeated GDPR right-to-be-forgotten. Archive now purges the FTS row inside the same SAVEPOINT.
• CLI --kind raw gated. Existing hippo forget / consolidation / conflict-resolution paths abort on raw rows via the trigger. Until those paths route through archiveRawMemory, the CLI restricts --kind to {distilled, superseded} so users cannot create unforgettable memories.
• NULL-kind trigger bypass. v14 triggers used WHEN NEW.kind IS NOT NULL AND NEW.kind NOT IN (...), so a direct kind=NULL write silently bypassed the CHECK substitute. v15 rejects NULL.
• archiveRawMemory transaction safety. Now uses SAVEPOINT (nestable) instead of BEGIN. BigInt-safe JSON serializer for the audit payload.
• --scope envelope trim. Matched the pre-existing scope-tag trim behavior.

Internal

• 730 tests (+15 from v0.33.0). New: tests/a3-envelope-migration.test.ts, tests/raw-archive.test.ts, tests/recall-why-envelope.test.ts.
• Reviewed via /codex, /plan-eng-review, /review (Claude pass + adversarial subagent), /self-review, /ship-check. All ship-blockers resolved before release.

v0.33.0 — Fact extraction + DAG summarization + multi-hop recall

0.33.0 (2026-04-23)

Added

• Write-time fact extraction. During hippo sleep, episodic memories are now processed by an LLM to extract standalone facts (up to 8 per memory). Facts are stored as semantic-layer entries with extracted_from linking back to the source. Extracted facts get a 1.3x search boost and automatically deduplicate against their source entries in results, so users see the precise fact instead of the raw conversation.
• DAG summarization. Extracted facts are clustered by Jaccard similarity (>= 0.5) on speaker:/topic: entity tags, then summarized into dag_level=2 parent nodes. When a summary matches a query, its children are injected into results at 0.9x parent score, giving hierarchical drill-down.
• Multi-hop retrieval. hippo recall --multihop and multihopSearch() run a two-pass entity-chained search. Pass 1 retrieves top-K and extracts entity tags not in the original query. Pass 2 reformulates the query with discovered entities and retrieves again. Results merge by highest score per ID.
• hippo remember --extract triggers immediate fact extraction on the remembered content.
• hippo dag --stats shows DAG layer distribution (how many entries at each level).
• Schema v12-v13. v12 adds extracted_from column, v13 adds dag_level + dag_parent_id with backfill and index. Backwards compatible, auto-migrates on first open.

Fixed

• temporalBoost O(N^2) refactored to O(N). Previously called Math.min(...timestamps) per entry inside the search loop, risking stack overflow on large stores. Now precomputes range once via computeTemporalRange().
• Config scoping bug in consolidate.ts. config was block-scoped inside the extraction if block but referenced from the DAG section outside it. Would cause ReferenceError when no extraction candidates exist but extracted facts are ready for DAG processing.
• Dead seenIds variables removed from both search paths (populated but never read).

Internal

• 674 tests (+41 from v0.32.0). 16 new test files covering extraction, DAG, multi-hop, temporal scoring, CLI commands, and integration smoke tests.
• Reviewed via /review + /self-review + /qa + /ship-check + senior code review agent.

v0.32.0 — Bi-temporal supersession + --as-of historical recall

0.32.0 (2026-04-22)

Added

• Bi-temporal memory: correction without deletion. When a belief changes, the old memory stays as historical truth instead of being overwritten. Default recall filters superseded entries so agents see current reality; historical views are explicit. Schema v11 adds valid_from and superseded_by columns, backwards compatible with v10 stores (ADD COLUMN only, no data transform).
• hippo supersede <old-id> "<new content>". Creates a successor memory and links the old one via superseded_by. Cycle prevention: if the target is already superseded, the command errors with the successor's ID so you can supersede that one instead. Reuses --layer, --tag, --pin from remember.
• --include-superseded on hippo recall / explain. Returns historical memories with a [superseded] marker in output. Default recall hides them.
• --as-of <ISO-date> on hippo recall / explain. Returns the set of memories that were current at that date. Validates input at CLI entry; invalid dates exit with a clear ISO-format hint.
• Partial index for fast current-only queries. CREATE INDEX idx_memories_current ON memories(layer, created) WHERE superseded_by IS NULL makes the default recall path cheap even with large archives.

Changed

• markRetrieved is a no-op for superseded entries. Retrieving a historical memory (via --include-superseded) no longer strengthens it or extends its half-life. Historical reads shouldn't revive dead beliefs.
• detectConflicts skips superseded pairs. No point flagging "these contradict" when one side is historically dead.

Research

• Physics search ablation: CUT verdict. Benchmarked physics-on vs physics-off over 60 stratified LongMemEval-oracle questions (paired bootstrap, 5000 iters). Physics OFF: MRR 0.8388, Recall@5 84.31%, NDCG@5 0.7888. Physics ON: MRR 0.6848, Recall@5 74.17%, NDCG@5 0.6570. All metrics statistically worse with physics; 95% CI excludes zero. Results in benchmarks/physics-ablation/. Physics remains in the codebase and is not removed in this release; a decision on removal is tracked as follow-up.
• LoCoMo harness built. benchmarks/locomo/run.py scores hippo against snap-research's long-conversation memory benchmark using Claude as judge. Sanity run (3 QAs): 2 adversarial abstentions correct, 1 open-domain miss. Full 10-conversation run requires overnight batch due to ~2 turns/sec ingestion.

Internal

• 633 tests pass (+8 from v0.31.0). 3 new test files: bi-temporal-migration.test.ts, cli-supersede.test.ts, bi-temporal-recall.test.ts.
• 4 commits on master: 091e6de (schema v11), 026988b (supersede command), b538c0d (recall filters), 7108187 (review fixes).
• Reviewed via /review + /self-review + /ship-check. Two fixes landed: --as-of date validation (previously silent no-op on invalid input) and cmdSupersede --tag parity with cmdRemember (previously only accepted comma-separated, dropped repeated flags).

v0.31.0 — Scope-aware corrections

Route corrections to the right context. Memories can now carry a scope tag, so a correction said during one skill surfaces strongly when that skill runs again, and stays quiet elsewhere.

Added

• hippo remember --scope <name> tags memories with a context scope.
• scopeBoost in search scoring — 1.5x when the active scope matches the memory's scope, 0.5x when it mismatches, 1.0x for unscoped. Same multiplier pattern as the existing decisionBoost / pathBoost / outcomeBoost.
• Auto-detect from env — HIPPO_SCOPE > GSTACK_SKILL > OPENCLAW_SKILL. When any is set, remember / recall / context / explain auto-apply the scope. Pure env var reads, no I/O on hot paths.
• --scope on hippo recall, hippo context, hippo explain — explicit scope overrides auto-detect.
• hippo explain --why surfaces the scope multiplier when it fires, so scope routing is debuggable.

Why

Inspired by claude-reflect's "skill improvement routing" — corrections during a specific skill should update that skill's context, not global memory. Hippo's take: same outcome via tag + scoring multiplier, no separate file hierarchy.

Internal

• 625 tests pass (+21 new, 0 regressions).
• New module src/scope.ts (32 lines). No schema migration. No breaking changes.
• Git-branch fallback was prototyped and dropped after /review — forked git on every hook call (~50-150ms per user message) and polluted the tag space with ephemeral branch names.

Upgrade

```bash
npm install -g hippo-memory@0.31.0
```

No migration. Additive only.

Full changelog

https://github.com/kitfunso/hippo-memory/blob/master/CHANGELOG.md#0310-2026-04-22

v0.30.1 — Sequence binding polish

Patch release: fixes three UX bugs found by code review + npm smoke test on v0.30.0. Sequence binding (the v0.30.0 headline feature) now works end-to-end as documented.

Fixed

• `hippo recall --layer ` is now a strict filter. Previously the flag was accepted but silently dropped, so results from other layers leaked in. This broke the RSI demo's `recall --layer trace` example.
• `hippo status` now prints a `Trace:` counter. The new trace layer was tracked internally but never surfaced.
• `hippo --version` / `-v` prints the package version. Previously errored with "Unknown command".

Internal

• 604 tests pass (+5 from v0.30.0). 3 new test files cover all three fixes via `execFileSync` against `bin/hippo.js`.
• Caught by `senior-code-reviewer` + npm smoke test against the published 0.30.0 artifact before the GitHub Release went public.

Upgrade

```bash
npm install -g hippo-memory@0.30.1
```

No migration. No breaking changes. If you're on v0.30.0, just upgrade.

v0.30.0 context

v0.30.0 introduced sequence binding — a new `Layer.Trace` memory type for ordered action→outcome sequences, foundation for recursive-self-improvement agents. See the v0.30.0 release notes for the full feature set.

Full changelog

https://github.com/kitfunso/hippo-memory/blob/master/CHANGELOG.md#0301-2026-04-22

v0.30.0 — Sequence binding (RSI foundation)

Sequence binding for recursive-self-improvement agents.

Agents can now remember step-by-step execution traces tagged with outcome (success / failure / partial), then recall them by outcome to learn from past attempts.

Highlights

• Layer.Trace — a new memory layer for ordered action→outcome sequences. Traces inherit decay, retrieval-strengthening, conflict detection, embeddings, replay, and physics from existing infrastructure. Four inheritance smoke tests lock that claim.
• hippo trace record --task <t> --steps <json> --outcome <...> — explicit trace storage.
• hippo session complete --session <id> --outcome <...> — terminal event marking a session as finished.
• hippo recall --outcome success — retrieve only successful prior strategies.
• Auto-promotion during hippo sleep — completed sessions become bound traces automatically. Idempotent; three consecutive sleeps produce exactly one trace per session.
• examples/rsi-demo/ — minimal RSI agent, deterministic and CI-runnable. 50-task suite. Current seed: 20% success on tasks 1-10 rising to 100% on tasks 41-50. Non-zero exit if the learning curve collapses.
• Schema v3 migration preserves existing data. 599 tests pass.

Known issues in this release (fixed in v0.30.1)

• `hippo recall --layer trace` was accepted but silently dropped, so results leaked other layers.
• `hippo status` had no `Trace:` counter.
• `hippo --version` / `-v` was missing.

Upgrade to v0.30.1 for the clean experience.

Full changelog

https://github.com/kitfunso/hippo-memory/blob/master/CHANGELOG.md#0300-2026-04-21--sequence-binding-recursive-self-improvement-foundation

v0.24.2 - machine-level daily runner + detached OpenClaw autosleep

What's new\n- Added a machine-level daily runner that sweeps registered Hippo workspaces and runs daily learn + sleep without requiring one OS task per project.\n- Updated the OpenClaw plugin to queue detached session-end autosleep so consolidation no longer blocks shutdown.\n- Updated the docs to separate query-time retrieval from session-end and daily refresh behavior across integrations.\n\n## Verification\n- npm run build\n- npm test\n- npm publish

v0.24.1 - conflict detection content-gate fix

Fixed

• Conflict detection now gates on content overlap, not shared tags. hippo sleep no longer flags unrelated feedback / policy memories as contradictions just because they share coarse tags and opposite polarity words.
• Reworded contradictions still surface. Opposites like API auth must be enabled in prod / Disable API auth in prod stay detectable instead of being filtered out by a blunt overlap threshold.
• must and always now count as positive polarity. Contradictions like Production deploys must require approval / Production deploys should not require approval are caught consistently.

Internal

• Added regression tests for the exact false-positive pairs from the migrated-store report plus a broader contradiction matrix (must vs should not, available vs missing, works vs broken). Full suite passes: 491 tests.