Show HN: Stop Claude Code from forgetting everything

How the tool works & intended use cases

• Skill connects Claude Code to an external MCP server that stores past conversations in a small DB (key/value + embeddings), organized by namespaces and “hypergraph” relationships.
• On each request it:
  • Embeds the current query.
  • Runs semantic + time-weighted search over prior sessions.
  • Returns only the top-N relevant snippets into the prompt as additional context.
• Used mainly to:
  • Resume long research/coding sessions across days.
  • Ask “what was I trying to do here?”, “what research threads already exist?”, “where did reasoning drift?”.
  • Let Claude reflect on and critique its own past reasoning.

Comparison to Claude’s built‑ins (CLAUDE.md, agents, skills, compaction)

• Several commenters say a good CLAUDE.md, AGENTS.md, per-project docs, and checkpoints/restore are enough; they see this as duplicating what agents + skills already solve.
• Others report:
  • Compaction making the model feel “dumber” and losing important edge cases.
  • CLAUDE.md often being ignored or only weakly applied.
• One thread explains a hierarchy:
  • CLAUDE.md → broad global/project instructions.
  • Agents → narrower, language/domain-specific instructions.
  • Skills → single-purpose instructions + deterministic tools (ripgrep, dependency graph analyzers, image generators), to keep context tight.

Privacy, hosting, and vendor lock‑in

• Multiple commenters say sending proprietary or sensitive code to a third‑party alpha service is a non‑starter; they want purely local or self‑hosted storage.
• Concerns include compliance, data leakage, vendor disappearance/price hikes, and negotiating agreements for “every small AI tool”.
• Some argue that even if useful, such features will eventually be best implemented by the model vendors themselves.

Alternatives and lightweight strategies

• Many describe simpler approaches:
  • Repo- or user-level CLAUDE.md and AGENTS.md.
  • Markdown “plans”, tickets, implementation logs, and work summaries committed to git.
  • Session JSONL parsing and local search (ripgrep, Tantivy, jq, custom CLIs).
  • Other memory tools: beads, claude-mem, Double, rg_history, memory-lane, custom MCP memory servers.
• Some find using less context, frequent fresh sessions, and strong planning/linting/tests more effective than elaborate memory layers.

Skepticism about memory abstractions & impact

• Repeated sentiment: there are already “countless” memory/context tools; few show benchmarks or clear productivity gains over simple docs.
• Doubts that external memory can reliably handle:
  • Drift, stale state, or subtle errors accumulating over time.
  • Multi-agent coordination without adding new failure modes.
• The project’s authors emphasize their focus on portability and shared state across tools/agents rather than “infinite context,” but some commenters remain unconvinced that semantic/temporal search alone solves the coordination problems they describe.