PiMem v1.4.0

PiMem v1.4.0 makes evidence retrieval and handoff more reliable while keeping the public search, read, and finish protocol stable.

• Adds lightweight optional working-memory progress notes and simpler finish correction handling.
• Preserves repeated-search sources and actual read receipts through context compaction and answer handoff.
• Splits search composition into smaller planning, discovery, execution, and candidate-set modules.
• Strengthens temporal, numeric, provenance, transport, concurrency, and Qdrant failure handling.
• Adds the official Fact-SH 262K and EventQA Full adapters required for the 1,600-query AMB-10 suite.

Validation: 453 TypeScript tests and 195 Python tests passed, together with type checking, documentation catalog validation, and a clean production build.

PiMem v1.3.1

PiMem v1.3.1 is a retrieval-boundary and reproducibility hardening release.\n\nHighlights:\n- Fall back from Qdrant to exact SQLite dense retrieval only for typed availability failures. Configuration, generation, scope, and provenance failures remain visible and fail closed.\n- Record dense fallback usage and result-affecting HNSW settings in retrieval metadata and formal runtime identities.\n- Pin the complete Qdrant configuration in protected MemoryAgentBench YAML, preventing ambient environment drift.\n- Separate SQLite lexical query planning from the persistence facade and require explicit dense-retriever injection in hybrid composition.\n- Strengthen bounded-context architecture tests and remove internal imports through public barrels.\n\nValidation:\n- TypeScript: 344 tests passed\n- Python integrations: 191 tests passed\n- Typecheck, generated architecture catalog verification, and production build passed

PiMem v1.3.0 — Hybrid Retrieval Baseline

PiMem v1.3.0 establishes the production hybrid-retrieval baseline.

Highlights:

• SQLite remains the immutable source of memory text, provenance, and lexical FTS5 retrieval.
• Qdrant HNSW provides the scalable dense lane; exact SQLite dense search is its failure fallback, not a duplicate RRF vote.
• Dense and lexical rankings are fused once with RRF, preserving complementary retrieval without double-counting dense evidence.
• Qdrant generations are immutable, resumable, fingerprinted, scope-verified, and revalidated against SQLite before evidence reaches the agent.
• MemoryAgentBench YAML runs now pin the Qdrant profile, collection, and generation in the runtime identity.

Validation:

• LME-S 300: 250/300 (83.33%) with gpt-5-mini medium for retrieval and answering, evaluated by the official GPT-4o judge; previous matched baseline: 236/300 (78.67%).
• BEAM-100K repeated runs: 74.21 and 73.34, mean 73.78.
• Full release check passed: TypeScript typecheck and build, 341 Vitest assertions, 66 MemoryAgentBench Python tests, and 107 MemoryArena Python tests.

Release commit: 1d83ca0.

PiMem v1.0.0

PiMem v1.0.0

PiMem 1.0 establishes the stable minimal contract for immutable, source-grounded Agent memory retrieval.

Highlights

• Agent-routed hybrid, lexical, coverage, temporal, numeric, and history operators through a minimal operator / queries / limit search API.
• Dense retrieval, SQLite FTS5, candidate-local BM25, and RRF fusion.
• Harness-owned scope enforcement, prepared SQL, query coverage, session aggregation, temporal/numeric fact joins, ranking, and provenance.
• Internally consistent Agent evidence packages retaining summaries, citation supports, counts, inventories, and cited raw memories.
• Exact immutable reads, durable traces, opaque harness references, and enforced Citations ⊆ Evidence ⊆ Candidates.
• Benchmark adapters remain responsible for final answer synthesis; ingest never accesses generative models or benchmark labels.

Validation

• TypeScript: 22 test files, 96 tests.
• Python: 2 frozen-evaluation tests.
• Full build and typecheck pass on Node 22.19+.
• Two fresh LongMemEval-S Direct runs: 356/500 and 348/500.
• Multi-session: 82/133 and 75/133.
• Zero provenance, ordering, or opaque-ID violations in both release-candidate runs.

The historical v0.1.0 release remains available as the preview baseline.

PiMem v0.1.0

Initial PiMem release: immutable raw-memory ingest, FTS5 and hybrid retrieval, provenance-enforced evidence selection, durable benchmark runs, and LongMemEval evaluation tooling.