memweaver

# MemWeaver — Memory Profiler > Your Agent reads your memory every day, but does it truly *understand* you? MemWeaver digs into your memory files — long-term memory (`MEMORY.md`) and daily logs — to uncover preferences, behavioral patterns, and **hidden traits you might not even be aware of**, then confirms findings through an interactive questionnaire and outputs a structured user profile (YAML). ## What Makes This Different | Existing tools | What they do | What MemWeaver does differently | |---|---|---| | Mem0 / Zep | Memory **retrieval** | Not retrieval — **understanding** | | SimpleMem / LightMem | Memory **compression** | Not compression — **insight mining** | | ai-persona-os | Give AI a persona | Opposite direction: discover **your** persona from memory | **Core value**: MemWeaver finds the gap between what you *say* you prefer and what you *actually do* — then asks you about it. ## Overview - **Input**: `MEMORY.md` (long-term memory) + recent daily logs (`memory/*.md`) - **Process**: LLM deep analysis → batch interactive questionnaire - **Output**: Structured user profile at `output/profile_YYYYMMDD.yaml` ## Workflow ### Step 1: Collect Memory ```bash cd {baseDir} && python3 scripts/collect_memory.py --days 14 ``` The script reads long-term memory and recent logs, outputs JSON to stdout. The Agent parses `content.long_term` and `content.daily_logs` fields from the JSON. **Note**: If `estimated_tokens` exceeds 8000, consider reducing the `--days` parameter. ### Step 2: LLM Deep Analysis The Agent analyzes collected memory in 3 sub-tasks: #### 2.1 Basic Profile Extraction Extract factual information from memory across these dimensions: | Dimension | What to extract | Confidence source | |---|---|---| | Identity | Role, tech stack, MBTI | Explicit statements → 1.0 | | Work patterns | Active projects, decision style, creation preference | Behavioral inference → 0.7-0.9 | | Interests | Professional interests, hobbies, depth of engagement | Topic frequency → 0.6-0.9 | | Communication | Response depth preference, format preference, dislikes | Interaction pattern → 0.7-0.85 | | Long-term goals | Career direction, product plans, values | Explicit statements → 0.9-1.0 | Tag each field with a `confidence` score. #### 2.2 Hidden Pattern Mining This is MemWeaver's most valuable part. The Agent specifically analyzes these 6 types of hidden patterns: 1. **Decision patterns**: What does the user lean toward when facing multiple options? (Analysis-driven vs intuition? Fast vs slow decisions?) 2. **Time & energy allocation**: Does actual energy distribution (from log frequency) match user's self-description? 3. **Overlooked interests**: Topics that appear repeatedly but the user hasn't formally tracked 4. **Statement vs behavior contradictions**: Are stated preferences inconsistent with actual actions? 5. **Emotion/energy triggers**: What scenarios make the user especially productive or resistant? 6. **Unlabeled skills**: Abilities the user demonstrates but hasn't self-recognized Each finding needs **evidence** (citing specific memory content) and **reasoning logic**. #### 2.3 Project Importance Re-evaluation List every project and idea recorded in MEMORY.md, provide reassessment: - Current status (active / paused / archived / concept) - Suggested importance (high / medium / low / shelved) - Assessment reasoning (frequency in logs, recent activity, user investment) - Questions to confirm with user (if uncertain) ### Step 3: Interactive Confirmation (Batch Questionnaire) Interact with the user in **batch mode**, similar to a personality test. Each question is based on the user's actual memory content, not just showing analysis conclusions. **Core design principles**: - **Push 5 questions per batch**, user answers at once (e.g., "1A 2C 3B 4D 5A"), Agent provides feedback then pushes next batch - Questions reference user's real memories as context - Provide options (A/B/C/D or open-ended), user can choose or free-form - **Type B (hidden insight) questions should be ≥50% of total** — this is MemWeaver's core value - Three question types interleaved, not strictly separated by rounds **Question Design Rules**: The Agent designs 10-15 questions based on Step 2 analysis. Three types: #### Type A: Scenario Recall (validate profile facts, ≤25%) Reconstruct a real scene from memory, let user choose the best description. ``` 📋 Q1. Your memory shows you did [specific behavior] on [specific date]. For you, this was more like: A. [option: engineering intuition / habit-driven] B. [option: lesson learned] C. [option: personality-driven] D. Other: ___ ``` #### Type B: Hidden Insight (core value, ≥50%) **This is MemWeaver's most important question type.** The Agent uses specific evidence from memory to point out contradictions or blind spots between user's "self-perception" and "actual behavior". **Methodology**: 1. Find user's **explicit statement** (e.g., "I prefer X") 2. Find **contradictory behavioral records** (e.g., logs show consistently doing Y) 3. Present the contradiction to user, guide explanation via options 4. Options should include: acknowledge contradiction, deny, offer new explanation, other ``` 📋 Q5. Your memory says "[user's explicit statement]". But logs show from [date A] to [date B] you've been consistently doing [contradictory behavior]. These two things: A. Don't contradict — [reasonable explanation] B. Actually contradict — my real preference differs from self-perception C. Depends on context — [conditional explanation] D. Other: ___ > 🔍 Your words say X, but your actions say Y ``` **Hidden insight mining directions** (look for these 6 types of clues in memory): 1. **Statement vs behavior contradictions**: Stated preferences inconsistent with actual actions 2. **Time allocation truth**: Log frequency/length reveals real energy distribution vs stated priorities 3. **Silence signals**: Topics in MEMORY that disappear from logs → possible priority drift 4. **Energy fingerprint**: Length differences across log types → reveals energy sources 5. **Choice patterns**: Consistent tendencies when user faces decisions 6. **Unlabeled skills**: Abilities demonstrated but not self-recognized #### Type C: Priority Trade-off (re-evaluate project importance, ≤25%) Create resource-constraint scenarios, force user to choose between projects, revealing true priorities. ``` 📋 Q10. If you could only advance 2 personal projects next month (work doesn't count), your memory mentions these: [project list from MEMORY.md] Which two? A. [Project1] + [Project2] B. [Project1] + [Project3] C. [Project2] + [Project3] D. Other combination: ___ ``` **Question count and batching**: - Total **10-15 questions**, pushed in **2-3 batches** (~5 per batch) - Batch 1 (5 questions): 1-2 warm-up Type A + 3 Type B (hidden insights) - Batch 2 (5 questions): 3-4 Type B + 1-2 Type C (priority trade-offs) - Optional Batch 3 (2-3 questions): follow-up questions based on unexpected findings from previous batches - After each batch, Agent waits for user to answer all at once **Answer processing**: - After user submits a batch (e.g., "1C 2B 3A 4B 5C"), Agent processes collectively - Give 1 brief feedback per question, noting profile inference - If an answer reveals new insight leads, add follow-up questions in next batch - All answers recorded internally as profile evidence, aggregated into Step 4 **Completion**: - After the last batch, Agent outputs a brief **profile summary** (like a personality test result page) - Then proceeds to Step 4 ### Step 4: Generate and Save Profile Generate the confirmed profile as YAML and save via script: 1. Agent generates complete YAML profile (see "Profile Template" below) 2. Save via script: ```bash cd {baseDir} && python3 scripts/save_profile.py --file /tmp/memweaver_profile.yaml ``` Or via stdin: ```bash echo '<YAML content>' | cd {baseDir} && python3 scripts/save_profile.py ``` The script automatically backs up old profiles and saves to `output/profile_YYYYMMDD.yaml`. --- ## Profile Template ```yaml # MemWeaver User Profile # Generated: YYYY-MM-DD # Version: 1 identity: role: "" tech_stack: [] mbti: "" confidence: 1.0 work_patterns: decision_style: "" # data_driven / intuitive / consultative detail_preference: "" # high / medium / low creation_preference: "" # 0to1 / polish / both energy_source: "" # ideation / execution / collaboration work_rhythm: "" # burst / steady / mixed confidence: 0.0 interests: professional: - topic: "" depth: "" # expert / active_research / exploring / casual importance: "" # high / medium / low last_active: "" personal: - topic: "" depth: "" importance: "" communication: preferred_depth: "" # deep_analysis / balanced / brief preferred_format: "" # structured / narrative / mixed language: "" dislikes: [] goals: career: - goal: "" priority: "" # high / medium / low / shelved timeframe: "" # immediate / short_term / long_term products: - name: "" priority: "" status: "" # active / designing / idea / archived hidden_patterns: - pattern: "" evidence: "" confirmed: false projects: - name: "" importance: "" # high / medium_high / medium / low / shelved status: "" # active / iterating / designing / exploring / archived / idea meta: generated_at: "" memory_files_analyzed: 0 total_memory_lines: 0 user_confirmed: true next_review: "" # Suggest re-profiling in 2 weeks ``` ## Output - User profile: `{baseDir}/output/profile_YYYYMMDD.yaml` - Analysis cache: `{baseDir}/cache/analysis_cache.json` (future version) ## Requirements No external dependencies. Python 3.8+ standard library only.