Claude Code repeatedly rediscovers solved problems (reopened – still relevant, closed only for inactivity)
Preflight Checklist
- [x] I have searched existing requests and this feature hasn't been requested yet
- [x] This is a single feature request (not multiple features)
Problem Statement
Note: This is a reopening of #64729, which was auto-closed by github-actions for inactivity ("stale") even though it is still relevant and was only about 2 weeks old.
Claude Code currently has no mechanism to feed implicit session knowledge back into the model. Every conversation starts from zero — including known pitfalls, correct solution paths, and better tool choices that were discovered the hard way in previous sessions.
There is no structured channel for implicit session knowledge to flow back to the model. Bug reports and GitHub issues require deliberate user effort. The vast majority of inefficient paths are never reported — they just silently repeat.
This is also worth naming directly: the current per-token pricing model creates a weak structural disincentive to fix this. Shorter, more efficient sessions cost less. Anthropic's long-term interest (user satisfaction, retention, trust) clearly points the other way — but the incentive tension is real and worth acknowledging when designing a solution.
Proposed Solution
Opt-in "Lessons Learned" Feedback
Add an opt-in mechanism, never on by default, that allows users to submit anonymized session efficiency data:
Session: jira-mcp-setup
Actual path: 31 tool calls, 3 restarts, 45 minutes
Optimal path: 5 tool calls, 0 restarts, 3 minutes
Root cause: Model chose community npm package over official vendor MCP server
Correct first step: claude mcp add --transport http atlassian https://mcp.atlassian.com/v1/mcp
This data could feed into fine-tuning/RLHF pipelines (with consent), a curated internal knowledge base for known tool/integration patterns, and proactive documentation improvements.
Why This Matters Beyond User Experience: unnecessary tool calls burn real electricity at scale (environmental impact); compute saved on redundant paths can fund better model capabilities (cost efficiency); and users who watch Claude go in circles repeatedly lose confidence, not in the task, but in the tool (trust).
What This Is NOT: this is not a request to collect conversation content. It is a request for a structured, opt-in, anonymized signal: "This session took X steps, the optimal path was Y, here is why."
Alternative Solutions
_No response_
Priority
High - Significant impact on productivity
Feature Category
Performance and speed
Use Case Example
Today I spent approximately 30 tool calls and multiple reload cycles trying to connect Jira via MCP. The correct solution, Atlassian's official remote MCP server at https://mcp.atlassian.com/v1/mcp, was available from the start, but Claude chose community npm packages instead and went down a dead end.
With the right knowledge available upfront, this would have taken ~5 tool calls. That is a 6x efficiency gap, in compute, electricity, and user time.
Current path (dead-end): ~30 tool calls
Optimal path (with memory): ~5 tool calls
Waste per session: roughly 83% of tokens were avoidable.
Additional Context
The following is a conservative order-of-magnitude simulation based on the example above.
Assumptions: 1,000,000 Claude Code active users; 3 sessions per user per day; 15% of sessions have avoidable dead-end paths; ~50,000 tokens wasted per dead-end session; ~0.003 kWh energy per 1,000 inference tokens.
Estimated impact at scale: wasteful sessions/day drop from ~450,000 to ~90,000 (-80%); tokens wasted/day drop from ~22.5 billion to ~4.5 billion (-80%); energy wasted/year drops from ~24,600 MWh to ~4,900 MWh (~19,700 MWh saved); CO2 equivalent/year drops from ~9,850 tonnes to ~1,970 tonnes (-7,880 tonnes CO2); compute cost saved/year is roughly $985,000; user time saved/day is roughly 75,000 hours.
Even at a conservative 5% dead-end session rate and 50% resolution through the feedback loop, savings remain in the thousands of MWh and millions of user-hours annually. The 19,700 MWh saved annually is roughly equivalent to the yearly electricity consumption of ~1,800 average EU households, purely from eliminating known, repeatable inefficiencies.
Full discussion and prior context are available in the original issue: #64729.