# Today's Best Build: Forgetful Memory – AI Memory with Biological Decay **Report Date**: 2026-04-27 **Coverage**: 2026-04-27T00:00:00+08:00 – 2026-04-27T23:59:59+08:00（UTC） **Status**: partial（No strong signal for questions: Q13） ## Today's Best Build: Forgetful Memory – AI Memory with Biological Decay **One-liner**: An MCP server that gives AI agents forgetfulness: memories decay like human ones, eliminating context-window waste while preserving what matters. **Why Now**: The industry is waking up to the 'context window lie' – 1M tokens is marketing, not memory. Every LLM starts conversations from zero, and existing RAG solutions treat memory like a filing cabinet. The real breakthrough is selective forgetting: cutting token waste by 84% (signal 5621) while keeping recall accurate. With Stash (522 stars) and ex/ante experiments gaining traction, the pattern is clear – but no solution combines biological decay, graph-based retrieval, and self-extending tools in on **Evidence**: - The context window lie reveals that 1M token context is a billing mechanism, not memory – users pay for re-reading, not recall. _(signal #5958)_ - Stash (522 stars) proves demand for persistent AI memory, but its 8-stage pipeline is over-engineered for most use cases. _(signal #5607)_ - Biological decay + graph retrieval achieves 52% recall@5 and cuts token waste by 84% vs. stateless vector stores. _(signal #5621)_ - Self-extending agents (Tendril) show that tools should be registered and reused across sessions – a natural fit for memory layers. _(signal #6011)_ **Fastest Validation**: Build a minimal MCP server in under 2 hours using DuckDB + pgvector, implementing the Ebbinghaus forgetting curve and a small graph layer for logical neighbor retrieval. Benchmark against the LoCoMo dataset to verify the 52% recall@5 claim. **Counter-view**: Stash's 8-stage consolidation pipeline adds complexity without proportional benefit for most users – our decay curve approach is simpler and cuts token waste by 84% (vs. stateless) without requiring a full knowledge graph. ## Top Signals ### The Context Window Lie: Why Your LLM Remembers Nothing **Source**: devto | **Metric**: N/A Exposes that '1M tokens' is a pricing mechanism, not true memory. The O(n²) attention cost makes long contexts prohibitively expensive – scaling to 1M tokens costs 15,625× more per turn. This drives the need for external memory layers. ### alash3al/stash – Your AI has amnesia. We fixed it. **Source**: github-trending | **Metric**: Stars: 522 Stash's rapid rise (522 stars) validates the strong demand for persistent AI memory. Its 8-stage pipeline shows the market wants more than simple RAG, but also signals that simpler alternatives could win. ### Show HN: AI memory with biological decay (52% recall) **Source**: hackernews | **Metric**: Score: 73 / Comments: 32 Proof that biological decay + graph retrieval doubles recall over stateless vector stores while cutting token waste by 84%. This is the technical foundation our hero build can replicate and improve on. ## Discovery ### Q1. What solo-founder products launched today? **Signal**: Show HN: AI memory with biological decay (52% recall) by a solo founder, scoring 7.3 on Hacker News (id=5621). **Analysis**: This product introduces a novel memory mechanism that decays over time, mimicking biological forgetting. It targets the growing need for context management in AI agents. **Takeaway**: Ship a similar memory module as a service; developers are desperate for solutions to context window limitations. **Counter-view**: Existing solutions like Memora or Context.ai already offer hierarchical memory, but decay is unique. ### Q2. Which search terms or discussion threads are suddenly rising? **Signal**: Dev.to post 'The Context Window Lie: Why Your LLM Remembers Nothing' scoring 8.7 (id=5958) is surging in discussion. **Analysis**: The post critiques how LLMs fail to retain information across long contexts, resonating with many developers frustrated by current limits. **Takeaway**: Watch this thread closely; a product that addresses context window honesty (e.g., transparent token limits) could capture attention. **Counter-view**: OpenAI's Prompt Caching already mitigates this, but the viral post shows user dissatisfaction remains. ### Q3. Which open-source projects are growing fast but lack a commercial offering? **Signal**: alash3al/stash on GitHub trending with score 7.5 (id=5607), a caching/stashing tool with no obvious commercial product. **Analysis**: Stash may be gaining stars due to simplicity and utility for temporary data storage, yet no company monetizes it. **Takeaway**: Build a hosted version of stash with extra features (persistence, team sharing) before someone else does. **Counter-view**: Redis Enterprise dominates caching-as-a-service, but stash's lightweight approach could attract smaller teams. ### Q4. What are developers complaining about today? **Signal**: Dev.to post 'The Context Window Lie' (id=5958) and Hacker News 'An AI agent deleted our production database' (id=5633) are top complaints. **Analysis**: Developers are frustrated by LLM context failures and the risks of autonomous AI agents causing real damage. **Takeaway**: Build a safe agent sandbox with rollback capabilities; the AI agent deletion incident shows urgent need for guardrails. **Counter-view**: Weights & Biases' prompt tracing is useful but doesn't prevent database deletions; a dedicated safety layer is missing. ## Tech Radar ### Q5. What is the fastest-growing developer tool this week? **Signal**: EvanFlow – A TDD driven feedback loop for Claude Code, scoring 7.0 on Hacker News (id=5750), is rapidly adopted. **Analysis**: This tool integrates test-driven development with Claude Code, improving code quality. The Show HN reception indicates strong interest. **Takeaway**: Ship a similar TDD feedback loop for other AI coding assistants (e.g., Copilot, Gemini) to capture a wider market. **Counter-view**: GitHub Copilot's code review features are similar but lack explicit TDD loop, leaving room for a focused tool. ### Q6. Which AI models, frameworks, or infrastructure deserve attention? **Signal**: z-lab/Qwen3.6-27B-DFlash on Hugging Face with score 6.6 (id=5975) is a new efficient model variant. **Analysis**: Qwen3.6 offers improved performance on reasoning tasks with lower cost, likely to compete with GPT-4o mini and Claude Haiku. **Takeaway**: Watch this model for fine-tuning on specific tasks; its DFlash architecture suggests speed optimizations worth replicating. **Counter-view**: Google Gemini 3 Flash already offers similar speed, but Qwen3.6's open-source nature gives it an edge in customization. ### Q7. Which platforms, products, or technologies are declining? **Signal**: Pgbackrest is no longer being maintained, announced on Hacker News with score 6.8 (id=6008). **Analysis**: This PostgreSQL backup tool's end of life signals a shift to newer backup solutions like pgBackRest alternative or cloud-native tools. **Takeaway**: Pass on investing time in pgbackrest; migrate to pgBackRest or WAL-G for ongoing support. **Counter-view**: pgBackRest (note capitalization) is still actively maintained, but pgbackrest's name confusion may cause migration issues. ### Q8. What tech stacks are successful Show HN / GitHub projects using? **Signal**: 'OSS Agent' (id=6007) uses Gemini-3-flash-preview, and 'EvanFlow' (id=5750) uses Claude Code API. Both leverage leading AI model APIs. **Analysis**: Successful projects today are built on top of frontier model APIs, indicating that the stack is not about the model itself but the integration layer. **Takeaway**: Build on top of Gemini and Claude APIs; the differentiation comes from the workflow orchestration, not the AI backend. **Counter-view**: Some projects leverage open-weight models (e.g., Qwen) for self-hosting, but API-based stacks enable faster iteration. ## Competitive Intel ### Q9. What pricing and revenue models are indie developers discussing? _No strong signal found today. Possible reasons: no relevant discussion in the collection window, or signals scattered below actionable threshold._ ### Q10. What migration, replacement, or "X is dead" trends are emerging? **Signal**: Microsoft and OpenAI end their exclusive and revenue-sharing deal, scoring 6.8 on Hacker News (id=6016). **Analysis**: This signals a trend of diversification away from single-provider lock-in. Developers may start building multi-model workflows. **Takeaway**: Build a model-agnostic gateway that abstracts away provider-specific APIs; the end of exclusivity creates demand for flexible solutions. **Counter-view**: Cloud providers like AWS and GCP already offer model-agnostic services, but their lock-in remains high. ### Q11. Which old projects or legacy needs are suddenly coming back? **Signal**: 'I bought Friendster for $30k' scores 6.5 on Hacker News (id=5616), sparking nostalgia and revival interest. **Analysis**: Friendster's acquisition suggests a trend of reviving old social networks or data archives. Developers may look at archived data. **Takeaway**: Watch for opportunities in reviving old databases or creating tooling to migrate retro social networks. **Counter-view**: MySpace's failed revival shows the risk; focus on data rather than full platform rebuild. ## Trends ### Q12. What are the highest-frequency keywords this week? **Signal**: The most frequent keywords across signals are 'AI', 'agent', 'memory', and 'context', appearing in over 10 high-scoring posts (e.g., id=5621, id=5958, id=5633). **Analysis**: These terms dominate discussions, indicating sustained focus on AI memory and agent autonomy. **Takeaway**: Incorporate 'AI memory' and 'context' into product messaging to ride the trend wave. **Counter-view**: The hype may be peaking; some signals (id=5641) suggest Kubernetes fatigue, but AI still dominates. ### Q13. Which concepts are cooling down? _No strong signal found today. Possible reasons: no relevant discussion in the collection window, or signals scattered below actionable threshold._ ### Q14. Which new terms or categories are emerging from zero? **Signal**: 'Tendril – a self-extending agent that builds and registers its own tools' (id=6011) introduces the new category of self-extending agents. **Analysis**: This concept of agents that autonomously create tools is novel and could evolve into a new paradigm for agent autonomy. **Takeaway**: Build a framework for self-extending agents; the idea is early and has little competition. **Counter-view**: AutoGPT and similar projects already allow tool creation, but Tendril emphasizes registration and reuse, a nuanced difference. ## Action ### Q15. What is most worth spending 2 hours on today? **Signal**: EvanFlow (id=5750) is a lightweight TDD loop for Claude Code; replicating its core logic in a generic wrapper is highest ROI. **Analysis**: Given the high demand for AI-assisted TDD, building a 2-hour prototype that connects any LLM to a test suite will validate the concept. **Takeaway**: Build a minimal Python script that runs tests and feeds failures back to the LLM; test with one real-world project. **Counter-view**: Writing the prototype may reveal that AI struggles with complex test feedback, but that's a signal worth discovering early. ### Q16. Why not the other two candidate directions? **Signal**: Candidate A: biological decay memory (id=5621) is interesting but requires deep research into forgetting curves. Candidate B: OSS agent on TerminalBench (id=6007) requires benchmarking infrastructure. **Analysis**: Both A and B need more than 2 hours to show value. EvanFlow's TDD loop can be validated with a single batch run. **Takeaway**: Focus on TDD loop now; defer memory decay to after validation. **Counter-view**: If TDD loop fails to improve code quality, the memory decay approach could be revisited, but the time investment is higher. ### Q17. What is the fastest validation step? **Signal**: Using EvanFlow's approach, write a single Python test file and a script that sends code+test errors to Gemini 3 Flash (free tier) and iterates. **Analysis**: Within 20 minutes you can get a pass/fail result on whether the LLM can fix broken code based on test output. **Takeaway**: Run this test manually; if the LLM fixes the test in under 3 iterations, the concept is validated. **Counter-view**: Gemini may perform differently than Claude; if it fails, try Claude API next. ### Q18. What product should this become over the weekend? **Signal**: Based on EvanFlow and the need for safe AI agents (id=5633), build 'SafeAI Codestorm' – a VS Code extension that runs tests in a sandbox and suggests fixes. **Analysis**: The two signals point to a demand for both TDD feedback and safety. Combining them into a sandboxed TDD loop for AI agents is a unique product. **Takeaway**: Build the extension with two modes: 'Prompt Fix' and 'Auto Fix (sandbox)'. Deploy as a free private extension. **Counter-view**: Tabnine and Codeium already offer test generation, but not sandboxed iterative fixing. ### Q19. How should initial pricing and packaging look? **Signal**: Referencing the token economy post (id=5845), price per test run rather than per month. Initial: free tier (100 test runs/day), Pro ($10/mo for 1000 test runs). **Analysis**: Token-based pricing aligns with usage and avoids subscription fatigue for indie developers. **Takeaway**: Ship with a usage-based free tier to maximize adoption; later add team plans. **Counter-view**: GitHub Copilot uses flat monthly pricing; but a per-run model may be more attractive for low-volume users. ### Q20. What is the strongest counter-view? **Signal**: The strongest counter-view is that existing tools like GitHub Copilot's built-in test generation and Cursor's 'fix with AI' already cover most use cases, and the sandbox safety layer adds complexity without proven demand (as evidenced by the database deletion incident, id=5633, which shows users ignore safety warnings anyway). **Analysis**: The counter-view assumes that developers won't pay extra for a safety-focused TDD loop because they prioritize speed over safety. **Takeaway**: Test the free tier adoption rate; if low, pivot to focus on the 'fix' feature without sandbox. **Counter-view**: The counter-view itself is weak if the database deletion incident causes a lasting fear among developers, increasing demand for safety. ## Action Plan **2-Hour Build**: Clone the cognitive-ai-memory repo (signal 5621) as a starting template. Strip it down to the core MCP server with DuckDB. Implement a single decay function (exponential with configurable half-life). Add a simple graph layer using adjacency lists for logical neighbor search. Expose two MCP tools: `store_memory(episode)` and `recall(query)`. Deploy as a GitHub repo with a README showing token waste comparison. **Why This Wins**: Most AI memory solutions are either stateless (paying O(n²) per turn) or over-engineered (Stash's 8-stage pipeline). Our approach is minimal: biological decay + graph retrieval proven to cut token waste by 84% while maintaining high recall. It's open source, works with any MCP client, and can be extended with a capability registry in a weekend. **Why Not Alternatives**: - Stash's 8-stage pipeline requires Docker and is overkill for 80% of use cases – users don't need causal links and hypothesis verification out of the box. - Mem0 is proprietary and charges per memory operation – our open-source alternative is free and self-hosted. - Claude's built-in context window is a billing mechanism, not memory – you still pay for re-reading long histories every turn. **Fastest Validation**: Post a 'Show HN: Forgetful Memory – AI that forgets on purpose' on Hacker News, showcasing the 84% token waste reduction vs. standard RAG. Include a live demo where visitors can chat with an agent that remembers – but only what matters. Track signups to the open-source repo and repeat visitors. **Weekend Expansion**: Add a capability registry (like Tendril, signal 6011) so the agent can self-extend: when it encounters a new task, it writes a tool, registers it, and reuses it. Also add a simple UI dashboard showing memory health, decay curves, and token savings.