Five Remedies for Claude Code's Amnesia: A Hands-On Comparison of Five Memory Solutions

> Problem: Claude Code forgets things — but "amnesia" is not one problem, it's five symptoms. > Solutions compared: Native capabilities / MemPalace / Hermes Agent / Superpowers / Everything Claude Code > Evaluation dimensions: Cold-start recovery, long-conversation fidelity, precise recall, cost

Key points

• "Amnesia" in Claude Code is actually five distinct symptoms, and misdiagnosis leads to wrong fixes:
1. Cross-session amnesia — new session, zero project context. 2. Context rot — quality drops past ~50% context usage; hallucinations spike past ~70%. 3. Precision recall gap — vague impressions but no exact retrieval of prior decisions. 4. Team knowledge silos — accumulated context isn't shared across teammates. 5. User modeling — the reverse need: Claude should remember *you*, not just the project.

Solution 1: Native Claude Code (free baseline)

Three native layers:

• CLAUDE.md: project rules loaded at startup; ~150–200 lines works best, adherence drops past 300.
• Auto Memory (v2.1.59+): Claude decides what's worth remembering into MEMORY.md; no search or structure — luck-dependent.
• /compact: manual compression that preserves CLAUDE.md but evaporates in-session decisions and constraints.

Ratings: cold-start ⭐⭐⭐, fidelity ⭐⭐, recall ⭐, cost ⭐⭐⭐⭐⭐ (free). Best for small personal projects with manual maintenance.

Solution 2: MemPalace (dedicated memory)

Open-source plugin treating memory as a first-class citizen:

• Two layers: a ~150-line CLAUDE.md plus a full memory store in .memory/state.json, searchable via MCP tools.
• Silent Hook capture: Stop (Haiku extracts Architecture / Decisions / Patterns / Gotchas / Progress / Context), PreCompact (rescues key info before compression), SessionEnd (final archive).
• Smart housekeeping: Jaccard deduplication, confidence decay (Progress 7 days, Context 30 days), LLM auto-merging, line budgets.

Ratings: cold-start ⭐⭐⭐⭐, fidelity ⭐⭐⭐⭐, recall ⭐⭐⭐⭐, cost ⭐⭐⭐. Best for precise recall and cross-session continuity; not for team sharing or user modeling.

Solution 3: Hermes Agent (user modeling)

A Nous Research project on cross-session long-term memory. Core idea: Claude should remember who the user is — preferences, decision patterns, behavioral profiles — via periodic reflection. Still research-grade, not production-ready.

Ratings: cold-start ⭐⭐⭐⭐, fidelity ⭐⭐⭐, recall ⭐⭐⭐, cost ⭐⭐ (heavy profiling overhead). For early adopters wanting "learns me over time"; not for production teams yet.

Solution 4: Superpowers (engineering discipline)

Jesse Vincent's methodology framework (TDD, YAGNI, structured debugging) that treats memory indirectly:

• Socratic brainstorming: clarify requirements before coding.
• 2–5 minute atomic tasks with fresh subagents — context stays at ~30–40% utilization, far from the 70% danger line.
• RED-GREEN-REFACTOR enforced TDD — untested code gets deleted.
• Plan documents enable precise audit of implementations.

Ratings: cold-start ⭐⭐⭐⭐⭐, fidelity ⭐⭐⭐⭐⭐, recall ⭐⭐⭐⭐, cost ⭐⭐⭐⭐. It doesn't make Claude remember more — it makes Claude need to remember less. Best for engineering teams; too slow for rapid prototyping.

Solution 5: Everything Claude Code (full harness)

ECC is a full-featured harness (Agent → Skill → Hook → Learning layers) where memory is a byproduct:

• SessionStart / Stop / SessionEnd / PreCompact / Pre- & PostToolUse hooks capture observations.
• Continuous Learning v1 (skill-based, ~50–80% coverage) and v2 (instinct-based, 100% hook coverage, confidence 0.3–0.9); /evolve merges 3+ instincts into reusable skills.

Ratings: cold-start ⭐⭐⭐, fidelity ⭐⭐⭐⭐, recall ⭐⭐⭐, cost ⭐⭐ (heavy token and maintenance overhead). Good as an upgrade if you already run a harness; overkill if you only want memory.

Spectrum positioning

| Solution | Positioning | Best symptom | |---|---|---| | Native | Free baseline | No budget, small projects | | MemPalace | Specialist | Precise recall, cross-session coherence | | Hermes | Research direction | User modeling (future) | | Superpowers | Engineering discipline | Long-conversation fidelity, cold start | | ECC | Full harness | Existing harness users |

Recommendations

• Everyone: master CLAUDE.md (150–200 lines), learn /compact, keep a hand-written project notes file.
• Need exact recall: adopt MemPalace; search history via MCP tools across sessions.
• Already using discipline frameworks: Superpowers/ECC come along for free.
• Want user modeling: watch Hermes / Nous Research — right direction, wrong timing.

What the community is really arguing about

The real debate is implicit vs. explicit memory: should the AI decide what to remember (Auto Memory), or should humans define memory structure (MemPalace, Superpowers)? The author's verdict: explicit wins today, because LLM self-summarization isn't reliable enough yet. Once models hit ~95% accuracy at judging what's worth remembering, implicit will win.

Anthropic is catching up; the third-party window may be short

Anthropic shipped Auto Memory (v2.1.59), expanded Hooks (8 event types), and the plugin ecosystem is exploding (176+ plugins, 20+ hooks). Once native structured memory + precise recall lands, plugins like MemPalace lose standalone value. But Superpowers and ECC endure — their value is methodology (TDD, YAGNI, Socratic brainstorming), which Anthropic won't choose for you.

Future: memory as infrastructure

• Short term (6 months): explicit memory plugins + discipline frameworks coexist.
• Mid term (1–2 years): native structured memory; third-party plugins move to vertical niches (user modeling, team sharing).
• Long term (3+ years): memory becomes infrastructure — you won't "install a memory plugin" any more than you install a "TCP/IP plugin."

The real competition isn't who remembers more — it's who remembers what matters.

References

• Claude Code official docs: memory.md, hooks.md
• MemPalace GitHub (yuvalsuede/memory-mcp)
• Hermes Agent (Nous Research)
• Superpowers (obra/superpowers) — Jesse Vincent
• Everything Claude Code (affaan-m/everything-claude-code) — Affaan Mustafa
• claude-mem (thedotmack/claude-mem); Claude-Recap (hatawong/claude-recap); claude-memory-engine (HelloRuru/claude-memory-engine); claude-hooks (mann1x/claude-hooks); cmr-memory (agynio/claude-map-reduce-memory); token-savior (Mibayy/token-savior)
• Hacker News discussion on the Claude Code plugin ecosystem (2026-04); Agentconn in-depth review