- Adjusted TTL durations and sleep intervals across memory and cache tests for consistent expiration behavior.
- Prevented test flakiness caused by timing discrepancies in token expiration and cache cleanup.
- Create shallow copy of VectorIndexEntry when adding similarity score
- Prevents mutation of cached entries that could corrupt shared state
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add _escape_like_pattern() helper to escape SQL wildcards (%, _, \)
- Apply escaping in SemanticMemory.search_facts and get_by_entity
- Apply escaping in ProceduralMemory.search and find_best_for_task
Prevents attackers from injecting SQL wildcard patterns through
user-controlled search terms.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add threading.Lock with double-check locking to ScopeManager
- Add asyncio.Lock with double-check locking to MemoryReflection
- Make reset_memory_metrics async with proper locking
- Update test fixtures to handle async reset functions
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Change source_episode_ids from JSON to JSONB for PostgreSQL consistency
- Add unique constraint for global facts (project_id IS NULL)
- Add CHECK constraint ensuring reinforcement_count >= 1
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Moved tests/unit/models/memory/ to tests/models/memory/ to avoid
Python import path conflicts when pytest collects all tests.
The conflict was caused by tests/models/ and tests/unit/models/ both
having __init__.py files, causing Python to confuse app.models.memory
imports.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Added `memory_consolidation` to the task list and updated `__all__` in test files.
- Updated `source_episode_ids` in `Fact` model to use JSON for cross-database compatibility.
- Revised related database migrations to use JSONB instead of ARRAY.
- Adjusted test concurrency in Makefile for improved test performance.
Auto-fixed linting errors and formatting issues:
- Removed unused imports (F401): pytest, Any, AnalysisType, MemoryType, OutcomeType
- Removed unused variable (F841): hooks variable in test
- Applied consistent formatting across memory service and test files
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add comprehensive caching layer for the Agent Memory System:
- HotMemoryCache: LRU cache for frequently accessed memories
- Python 3.12 type parameter syntax
- Thread-safe operations with RLock
- TTL-based expiration
- Access count tracking for hot memory identification
- Scoped invalidation by type, scope, or pattern
- EmbeddingCache: Cache embeddings by content hash
- Content-hash based deduplication
- Optional Redis backing for persistence
- LRU eviction with configurable max size
- CachedEmbeddingGenerator wrapper for transparent caching
- CacheManager: Unified cache management
- Coordinates hot cache, embedding cache, and retrieval cache
- Centralized invalidation across all caches
- Aggregated statistics and hit rate tracking
- Automatic cleanup scheduling
- Cache warmup support
Performance targets:
- Cache hit rate > 80% for hot memories
- Cache operations < 1ms (memory), < 5ms (Redis)
83 new tests with comprehensive coverage.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add MCP-compatible tools that expose memory operations to agents:
Tools implemented:
- remember: Store data in working, episodic, semantic, or procedural memory
- recall: Retrieve memories by query across multiple memory types
- forget: Delete specific keys or bulk delete by pattern
- reflect: Analyze patterns in recent episodes (success/failure factors)
- get_memory_stats: Return usage statistics and breakdowns
- search_procedures: Find procedures matching trigger patterns
- record_outcome: Record task outcomes and update procedure success rates
Key components:
- tools.py: Pydantic schemas for tool argument validation with comprehensive
field constraints (importance 0-1, TTL limits, limit ranges)
- service.py: MemoryToolService coordinating memory type operations with
proper scoping via ToolContext (project_id, agent_instance_id, session_id)
- Lazy initialization of memory services (WorkingMemory, EpisodicMemory,
SemanticMemory, ProceduralMemory)
Test coverage:
- 60 tests covering tool definitions, argument validation, and service
execution paths
- Mock-based tests for all memory type interactions
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add MemoryConsolidationService with Working→Episodic→Semantic/Procedural transfer
- Add Celery tasks for session and nightly consolidation
- Implement memory pruning with importance-based retention
- Add comprehensive test suite (32 tests)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add comprehensive indexing and retrieval system for memory search:
- VectorIndex for semantic similarity search using cosine similarity
- TemporalIndex for time-based queries with range and recency support
- EntityIndex for entity-based lookups with multi-entity intersection
- OutcomeIndex for success/failure filtering on episodes
- MemoryIndexer as unified interface for all index types
- RetrievalEngine with hybrid search combining all indices
- RelevanceScorer for multi-signal relevance scoring
- RetrievalCache for LRU caching of search results
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add scope management system for hierarchical memory access:
- ScopeManager with hierarchy: Global → Project → Agent Type → Agent Instance → Session
- ScopePolicy for access control (read, write, inherit permissions)
- ScopeResolver for resolving queries across scope hierarchies with inheritance
- ScopeFilter for filtering scopes by type, project, or agent
- Access control enforcement with parent scope visibility
- Deduplication support during resolution across scopes
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implements procedural memory for learned skills and procedures:
Core functionality:
- ProceduralMemory class for procedure storage/retrieval
- record_procedure with duplicate detection and step merging
- find_matching for context-based procedure search
- record_outcome for success/failure tracking
- get_best_procedure for finding highest success rate
- update_steps for procedure refinement
Supporting modules:
- ProcedureMatcher: Keyword-based procedure matching
- MatchResult/MatchContext: Matching result types
- Success rate weighting in match scoring
Test coverage:
- 43 unit tests covering all modules
- matching.py: 97% coverage
- memory.py: 86% coverage
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implements the episodic memory service for storing and retrieving
agent task execution experiences. This enables learning from past
successes and failures.
Components:
- EpisodicMemory: Main service class combining recording and retrieval
- EpisodeRecorder: Handles episode creation, importance scoring
- EpisodeRetriever: Multiple retrieval strategies (recency, semantic,
outcome, importance, task type)
Key features:
- Records task completions with context, actions, outcomes
- Calculates importance scores based on outcome, duration, lessons
- Semantic search with fallback to recency when embeddings unavailable
- Full CRUD operations with statistics and summarization
- Comprehensive unit tests (50 tests, all passing)
Closes#90🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implements session-scoped ephemeral memory with:
Storage Backends:
- InMemoryStorage: Thread-safe fallback with TTL support and capacity limits
- RedisStorage: Primary storage with connection pooling and JSON serialization
- Auto-fallback from Redis to in-memory when unavailable
WorkingMemory Class:
- Key-value storage with TTL and reserved key protection
- Task state tracking with progress updates
- Scratchpad for reasoning steps with timestamps
- Checkpoint/snapshot support for recovery
- Factory methods for auto-configured storage
Tests:
- 55 unit tests covering all functionality
- Tests for basic ops, TTL, capacity, concurrency
- Tests for task state, scratchpad, checkpoints
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implements Sub-Issue #87 of Issue #62 (Agent Memory System).
Core infrastructure:
- memory/types.py: Type definitions for all memory types (Working, Episodic,
Semantic, Procedural) with enums for MemoryType, ScopeLevel, Outcome
- memory/config.py: MemorySettings with MEM_ env prefix, thread-safe singleton
- memory/exceptions.py: Comprehensive exception hierarchy for memory operations
- memory/manager.py: MemoryManager facade with placeholder methods
Directory structure:
- working/: Working memory (Redis/in-memory) - to be implemented in #89
- episodic/: Episodic memory (experiences) - to be implemented in #90
- semantic/: Semantic memory (facts) - to be implemented in #91
- procedural/: Procedural memory (skills) - to be implemented in #92
- scoping/: Scope management - to be implemented in #93
- indexing/: Vector indexing - to be implemented in #94
- consolidation/: Memory consolidation - to be implemented in #95
Tests: 71 unit tests for config, types, and exceptions
Docs: Comprehensive implementation plan at docs/architecture/memory-system-plan.md
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Added detailed unit tests for `/context` endpoints, covering health checks, context assembly, token counting, budget retrieval, and cache invalidation.
- Included edge cases, error handling, and input validation for context-related operations.
- Improved test coverage for the Context Management module with mocked dependencies and integration scenarios.
- Introduced end-to-end tests for MCP workflows, including server discovery, authentication, context engine operations, error handling, and input validation.
- Added full lifecycle tests for agent workflows, covering type management, instance spawning, status transitions, and admin-only operations.
- Enhanced test coverage for real-world MCP and Agent scenarios across PostgreSQL and async environments.
- Introduced a new `context` module and its endpoints for Context Management.
- Added `/context` route to the API router for assembling LLM context, token counting, budget management, and cache invalidation.
- Implemented health checks, context assembly, token counting, and caching operations in the Context Management Engine.
- Included schemas for request/response models and tightened error handling for context-related operations.
- Introduced integration tests covering backend, LLM Gateway, Knowledge Base, and Context Engine.
- Includes health checks, tool listing, token counting, and end-to-end MCP flows.
- Added `RUN_INTEGRATION_TESTS` environment flag to enable selective test execution.
- Includes a quick health check script to verify service availability before running tests.
- Added stricter budget validation in ContextRanker with explicit error handling for invalid configurations.
- Introduced `_get_valid_token_count()` helper to validate and safeguard token counts.
- Enhanced XML escaping in Claude adapter to prevent injection risks from scores and unhandled content.
- Added timeout enforcement for token counting, scoring, and compression with detailed error handling.
- Introduced tenant isolation in context caching using project and agent identifiers.
- Enhanced budget management with stricter checks for critical context overspending and buffer limitations.
- Optimized per-context locking with cleanup to prevent memory leaks in concurrent environments.
- Updated default assembly timeout settings for improved performance and reliability.
- Improved XML escaping in Claude adapter for safety against injection attacks.
- Standardized token estimation using model-specific ratios.
- Cleaned up unnecessary comments in `__all__` definitions for better readability.
- Adjusted indentation and formatting across modules for improved clarity (e.g., long lines, logical grouping).
- Simplified conditional expressions and inline comments for context scoring and ranking.
- Replaced some hard-coded values with type-safe annotations (e.g., `ClassVar`).
- Removed unused imports and ensured consistent usage across test files.
- Updated `test_score_not_cached_on_context` to clarify caching behavior.
- Improved truncation strategy logic and marker handling.
- Add tests for truncation edge cases, including zero tokens, short content, and marker handling.
- Add concurrency tests for scoring to verify per-context locking and handling of multiple contexts.
- Replace hard-coded limits with configurable settings (e.g., cache memory size, truncation strategy, relevance settings).
- Optimize parallel execution in token counting, scoring, and reranking for source diversity.
- Improve caching logic:
- Add per-context locks for safe parallel scoring.
- Reuse precomputed fingerprints for cache efficiency.
- Make truncation, scoring, and ranker behaviors fully configurable via settings.
- Add support for middle truncation, context hash-based hashing, and dynamic token limiting.
- Refactor methods for scalability and better error handling.
Tests: Updated all affected components with additional test cases.
Phase 7 of Context Management Engine - Main Engine:
- Add ContextEngine as main orchestration class
- Integrate all components: calculator, scorer, ranker, compressor, cache
- Add high-level assemble_context() API with:
- System prompt support
- Task description support
- Knowledge Base integration via MCP
- Conversation history conversion
- Tool results conversion
- Custom contexts support
- Add helper methods:
- get_budget_for_model()
- count_tokens() with caching
- invalidate_cache()
- get_stats()
- Add create_context_engine() factory function
Tests: 26 new tests, 311 total context tests passing
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Phase 5 of Context Management Engine - Model Adapters:
- Add ModelAdapter abstract base class with model matching
- Add DefaultAdapter for unknown models (plain text)
- Add ClaudeAdapter with XML-based formatting:
- <system_instructions> for system context
- <reference_documents>/<document> for knowledge
- <conversation_history>/<message> for chat
- <tool_results>/<tool_result> for tool outputs
- XML escaping for special characters
- Add OpenAIAdapter with markdown formatting:
- ## headers for sections
- ### Source headers for documents
- **ROLE** bold labels for conversation
- Code blocks for tool outputs
- Add get_adapter() factory function for model selection
Tests: 33 new tests, 256 total context tests passing
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add TokenCalculator with LLM Gateway integration for accurate token
counting with in-memory caching and fallback character-based estimation.
Implement TokenBudget for tracking allocations per context type with
budget enforcement, and BudgetAllocator for creating budgets based on
model context window sizes.
- TokenCalculator: MCP integration, caching, model-specific ratios
- TokenBudget: allocation tracking, can_fit/allocate/deallocate/reset
- BudgetAllocator: model context sizes, budget creation and adjustment
- 35 comprehensive tests covering all budget functionality
Part of #61 - Context Management Engine
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Adjusted `docker-compose.dev.yml` to reflect the new venv location.
- Modified entrypoint script and Dockerfile to reference `/opt/venv` for isolated dependencies.
- Improved bind mount setup to prevent venv overwrites during development.
- Added `record_action` in `RateLimiter` for precise tracking of slot consumption post-validation.
- Introduced deduplication mechanism for warning alerts in `CostController` to prevent spamming.
- Refactored `CostController`'s session and daily budget alert handling for improved clarity.
- Implemented test suites for `CostController` and `SafetyGuardian` to validate changes.
- Expanded integration testing to cover deduplication, validation, and loop detection edge cases.
Improved code readability and uniformity by standardizing line breaks, indentation, and inline conditions across safety-related services, models, and tests, including content filters, validation rules, and emergency controls.
The delay2 and delay3 variables were calculated but never asserted,
causing lint warnings. Added assertions to verify all delays are
positive and within max bounds.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The ContentFilter was appending references to DEFAULT_PATTERNS objects,
so when tests modified patterns (e.g., disabling them), those changes
persisted across test runs. Use dataclass replace() to create copies.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add MCPSafetyWrapper for safe MCP tool execution
- Add MCPToolCall/MCPToolResult models for MCP interactions
- Add SafeToolExecutor context manager
- Add SafetyMetrics collector with Prometheus export support
- Track validations, approvals, rate limits, budgets, and more
- Support for counters, gauges, and histograms
Issue #63🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
