Show HN: 벤치마크를 통해 연결된 사실에 대한 AI 메모리 사각지대 수정

Three external datasets. All scripts are public and reproducible — full methodology in BENCHMARKS.md. snap-research/LoCoMo · 1,534 QA pairs · 10 sessions benchmarks/locomo_4way.py · BM25 + vector + graph + decay · 20 Apr 2026 Dataset Both systems ran all 10 samples to completion on the same 1,534 QA pairs. YourMemory led every single session — the only memory layer tested that completed the full benchmark clean. All retrieval, pruning, and graph expansion runs fully on your machine — no cloud inference cost, no data leaving your environment. 3-session developer workflow simulation — stateless baseline vs YourMemory −84% At 30 sessions. Memory block stays flat (~76–91 tokens) while stateless history grows O(n). At 3 sessions: −19.7% tokens, −28% per-session context. −14% Recalled context eliminates clarifying questions at the start of new sessions. Each clarifying round is a full LLM call that produces zero implementation output. −4% Memories below Ebbinghaus strength 0.05 are pruned from retrieval entirely. 3/15 memories pruned in a 60-day synthetic set. Compounds at scale (200+ memories). Vector search finds what you asked for. The graph finds what you forgot to ask for. Ebbinghaus decides what survives. Different kinds of memory age at different rates. Important facts persist longer; transient context fades naturally. Related memories stay alive together — no orphaned facts. Two-round retrieval finds not just what you searched for, but what you forgot to search for. Related memories surface even when they don't share vocabulary with the query. Multiple agents share context or keep secrets. API keys (ym_ prefix) authenticate each agent. Shared vs private visibility per memory. Semantic search alone misses memories that are related but worded differently. A second retrieval pass surfaces them automatically. Finds the most relevant memories for your query — fast and precise. Related memories that didn't match the query directly are surfaced through the graph layer — nothing slips through. Using a memory keeps its connected context cluster fresh automatically — the more you use it, the longer it survives. Connected memory graph Memories don't decay in isolation. Before a memory is pruned, its connected neighbours are checked — if any are still relevant, the whole cluster stays alive. Related facts age together. Every time a memory is recalled, its connected neighbours get a freshness boost. The more a cluster of related memories is used, the longer the whole group persists — the system learns what matters to you. Multiple AI agents share context or keep secrets. Each agent authenticates with an API key. You control exactly what each agent can read and write. Each agent gets a unique ym_ API key. Shown once, never stored in plaintext. Revoke anytime. Pass the API key in any MCP call. Set visibility to control who can see it. Without a key → shared memories only. With a key → shared + that agent's private memories. | Memory stored as | Owner agent | Other agents | No API key | |---|---|---|---| shared | ✓ | ✓ | ✓ | private | ✓ | ✗ | ✗ | Keys hashed with SHA-256 before storage. Revoke anytime with revoke_agent(agent_id, user_id) . yourmemory-setup automatically injects a curated instruction set into your agent's global context — telling it exactly when to recall, what to store, and how to prioritise memories. No manual configuration needed. Recall policy — agent retrieves context before every task automatically Store / update / ignore decision logic — no duplicate memories, no noise Importance and category guidance — agent assigns decay rates and priority without being told Written to ~/.claude/CLAUDE.md — applies globally across all your projects [1/4] Downloading spaCy model… ✓ en_core_web_sm installed [2/4] Initialising database… ✓ Database ready [3/4] Writing MCP config… ✓ Claude Code → ~/.claude/settings.json [4/4] Injecting memory rules… ✓ Memory rules → ~/.claude/CLAUDE.md ✓ Setup complete. Restart your AI client. Install, run setup. That's it — spaCy model, database, and client configs are handled automatically. $ pip install yourmemory $ yourmemory-setup Configures everything automatically — language model, database, and MCP config for every detected client on your machine. ✓ Language model ready ✓ Database initialised ✓ Claude Code → ~/.claude/settings.json ✓ Claude Desktop → auto-detected if installed ✓ Cursor / Windsurf / Cline → auto-detected if installed ✓ Memory rules → injected into global agent context Restart your AI client after setup. YourMemory starts automatically as an MCP server — no background process to manage. Works with PostgreSQL (optional — teams / large datasets) $ pip install yourmemory[postgres] DATABASE_URL=postgresql://YOUR_USER@localhost:5432/yourmemory Backend selected automatically from the connection string — no additional config required. $ pip install 'yourmemory[neo4j]' $ GRAPH_BACKEND=neo4j yourmemory Default graph runs fully in-process with zero setup. Switch to the production backend for large deployments via the GRAPH_BACKEND env var. Read the thinking behind it A deep dive into the architecture — why flat vector search isn't enough and how Ebbinghaus decay + graph traversal changes retrieval quality. The original writeup — the forgetting curve formula, category-based decay rates, and why importance modulates λ instead of just being a filter. Community discussion — the architecture, tradeoffs, and why Ebbinghaus decay outperforms flat vector stores. Launch post — the story behind building a memory layer that works like human memory, and the benchmark results from the public LoCoMo dataset. Coverage of how YourMemory applies the Ebbinghaus forgetting curve to AI agent memory — making context decay naturally so only relevant knowledge survives.