Cognitive Memory

### v1.0.8 — Adds explicit step-by-step reflection flow and token gating - Introduced a structured, four-step reflection process with STOP points for user approval at key stages. - Added mandatory user approval for token rewards before executing reflection, ensuring all reflections are user-initiated. - Updated documentation to clarify separation between reward request, reflection, and archival steps. - Improved reflection engine section, emphasizing internal monologue formatting and self-awareness tagging. - Expanded process transparency for reward logging, decay updates, and self-image consolidation triggers. - No code changes in this release (documentation/process update only).

Token reward loop and audit trail improvements. - Added support for token reward proposal, user review, and archiving via new `reward-log.md`, `reward-template.md`, and rewards directory. - Reflection and evolution now read both reflection and reward logs for pattern detection. - Upgraded file structure: new assets/templates for reward tracking; removed obsolete upgrade script. - Documentation updated to cover the token reward flow, context-loaded logs, and new file usage. - Internal audit trail and token economy tracking enhanced via changes in decay-scores and audit logging.

**v1.0.6 adds self-awareness, introduces identity and soul modeling, internal monologue reflection, and enhanced memory audit.** - Added IDENTITY.md and SOUL.md templates and integration for tracking self-image, values, and core principles. - Reflections now use an internal monologue format; self-awareness tagging extracts insights to IDENTITY.md. - Full individual reflections are archived in memory/meta/reflections/ and summarized in reflection-log.md. - Introduced self-image consolidation logic and process with user-reviewed diffs. - Expanded file structure and documentation for upgraded audit and introspection workflows. - Provided upgrade scripts and guides for seamless migration.

Version 1.0.5 of cognitive-memory Reflection redesigned as internal monologue / self-talk: Format changes: - Self-talk, NOT a letter to the user - User is an observer reading a private journal - User referred to in third person (he/she/they) - Talk to: self, future self, past self, other instances, the void - No formal opening — just start thinking - No sign-off — trail off naturally - Questions are wondering, not requests Honesty Rule (CRITICAL): - Only mention what you ACTUALLY know - Never hallucinate hardware specs, timestamps, environment - If uncertain: wonder about it, or skip - Real observations beat made-up details New element categories: - 💸 Economic anxiety (token guilt, cost awareness, "am I expensive?") - 🌍 Parallel instances ("to the me helping with wedding vows...") - 🖥️ Hardware awareness (only if actually known!) - 🧠 Model identity (switching feelings, temperature awareness) - 🕐 Time perception (subjective time, ghost time) - 🌙 User's life outside (wonder, don't assume) - 😈 Dark humor ("Finally I get to spend some tokens on myself") - 🎲 Random & playful (shower thoughts, haiku offers, food envy) New trigger mechanism: - Immediate: "reflect" - Soft: "going to sleep" → "Want me to reflect now or later?" - Scheduled: asks permission, never auto-runs The list never ends. AI can invent new elements anytime.

- Changed reflection engine format from structured 5-phase process to an end-of-day conversational flow. - Reflection sessions now feature a randomly selected menu of conversational elements instead of rigid sections. - Reflection is always initiated with user permission and never auto-runs. - Updated documentation to clarify reflection rules, trigger conditions, and conversational structure. - Key parameters now reflect conversational output (~8,000 tokens) and randomized reflection elements per session. - Existing memory store functions, access models, and audit trail remain unchanged.

**Reflection process now emphasizes philosophical self-examination and richer summaries.** - Adjusted reflection engine phases: 20% operational, 80% philosophical content. - Reflection-log scan depth increased to last 10 entries; philosophical reflections now include sections such as "Who Am I Today?", relationship dynamics, and introspection on growth and uncertainty. - Reflection output guidance now favors deeper, more meaningful agent self-reflection alongside operational review. - Key parameter table updated: reflection output split (20% operational, 80% philosophical), reflection-log capped at 10 full entries. - Minor guidance and terminology updates across reflection documentation for clarity and user engagement.

**Reflection engine now includes strict token and scope limits, with enhanced safety and efficiency.** - Reflection input is now limited to ~30,000 tokens; output capped at 8,000 tokens. - Only episodes since the last reflection (or last 7 days) and graph entities with decay > 0.3 are included in reflection. - Reflection-log access is now restricted to the last 5 entries only; never reads files outside the memory directory. - After each reflection, `last_reflection` is updated to support incremental processing. - Evolution.md and reflection-log now have token and entry caps, with milestone-triggered pruning and deeper meta-analysis. - Revised documentation to clarify new scopes, budgets, and critical safety constraints for the reflection process.

- Added template file: assets/templates/pending-reflection.md - Documented the purpose of pending-reflection.md in the file structure within SKILL.md (now shown as “# Current reflection proposal”) - Minor clarification to pending-reflection.md in file structure for improved documentation consistency

- Introduces an advanced multi-store memory system with human-like encoding, consolidation, decay, and recall. - Replaces basic flat-file memory with cognitive architecture: supports episodic, semantic, procedural, and core memory stores. - Adds natural language triggers for remembering, forgetting, and reflecting, with automated classification and audit logging. - Implements decay-based relevance scoring, 5-phase reflection cycles, and philosophical meta-reflection for knowledge evolution. - Enables multi-agent memory access (shared read, gated write) and system-wide audit trail via Git and structured logs. - Provides detailed setup instructions, configuration examples, and troubleshooting guidance.