[upstream PR 4] feat: agentmemory v0.3.0 — semantic search, memory evolution, profiles

[wbugitlab1]

Source repository: rohitg00/agentmemory
Source pull request number: 4
Source URL: intentionally omitted to avoid GitHub cross-references
Title: feat: agentmemory v0.3.0 — semantic search, memory evolution, profiles
Author: rohitg00
State: closed
Draft: no
Merged: yes
Head: rohitg00/agentmemory:feat/v0.3.0-competitive-upgrade @ 31cd297
Base: main @ 1e4f1f3
Labels: (none)
Changed files: 0
Commits: 0
Created: 2026-02-26T17:25:36Z
Updated: 2026-02-27T05:05:29Z
Closed: 2026-02-27T05:05:26Z
Merged at: 2026-02-27T05:05:26Z

Original PR body:

Summary

• Hybrid BM25 + vector semantic search with Reciprocal Rank Fusion (6 embedding providers: Gemini, OpenAI, Voyage AI, Cohere, OpenRouter, local @xenova/transformers)
• Memory versioning & relationships — supersedes, extends, derives, contradicts, related; BFS graph traversal
• Persistent search indices — debounced KV writes for both BM25 and vector indices; survives restarts
• Progressive disclosure search — compact results first (id, title, score), expand specific IDs for full details
• Project profiles — aggregated top concepts, files, conventions, common errors with 1-hour cache
• Chronological timeline — navigate observations around anchor points (ISO dates or keywords)
• Auto-forget — TTL expiry via forgetAfter field + Jaccard contradiction detection (threshold 0.9)
• Provider fallback chain — ordered LLM provider list with automatic failover
• Full JSON export/import — round-trip with merge/replace/skip strategies
• 10 MCP tools, 28 REST endpoints, 21 functions (up from 5/20/13 in v0.2.0)

Changes

Category	Files	Details
New source	17	6 embedding providers, vector index, hybrid search, index persistence, fallback chain, relations, timeline, smart-search, profile, auto-forget, export-import
New tests	11	62 new tests (144 total passing)
Modified	14	types, config, schema, search-index, providers, compress, remember, consolidate, context, evict, index, mcp/server, triggers/api, telemetry, eval/schemas
Build	39.88 KB	0 type errors

v0.2.0 → v0.3.0 comparison

Capability	v0.2.0	v0.3.0
Search	BM25 keyword only	Hybrid BM25 + vector (RRF)
Embeddings	None	6 providers (auto-detect from env)
Memory model	Flat KV	Versioned with relationships
Index persistence	In-memory (lost on restart)	KV-persisted with debounced writes
Search results	Full observations	Progressive disclosure
Profiles	None	Project profiles
Timeline	None	Chronological navigation
Forgetting	Manual eviction only	Auto-forget (TTL + contradiction)
Provider resilience	Circuit breaker	Circuit breaker + fallback chain
Data portability	SQLite import	Full JSON export/import
MCP tools	5	10
API endpoints	20	28
Functions	13	21

Test plan

• npm run build — 0 type errors, 39.88 KB bundle
• npm test — 144 tests pass, 14 integration skipped (needs running server)
• Start with GEMINI_API_KEY → logs "Embedding provider: gemini (768 dims)"
• Start without embedding keys → logs "Embedding provider: none (BM25-only mode)"
• Restart → logs "Loaded persisted BM25 index (N docs)"
• POST /agentmemory/smart-search returns compact results
• POST /agentmemory/timeline returns chronological observations
• GET /agentmemory/profile?project=... returns aggregated profile
• GET /agentmemory/export → POST /agentmemory/import round-trip
• POST /agentmemory/auto-forget?dryRun=true shows candidates

Summary by CodeRabbit

• New Features

  • Hybrid BM25+vector search with persistent index, multiple embedding providers (OpenAI, Cohere, Gemini, Voyage, OpenRouter, local), and fallback chains.
  • Memory versioning, relations, evolve, automatic cleanup (auto-forget/evict), smart search (compact/expand), timeline browsing, project profiling, and export/import (merge/replace/skip).
  • New HTTP and MCP endpoints to access these capabilities.

• Chores

  • Bumped release to v0.3.0; added telemetry metrics and updated dependencies.

Local branch:
Fork PR:
Fork decision:
Verification:
Notes: