\n Enterprise-grade, privacy-first Long-Term Memory (LTM) engine for LLM agents, multi-agent frameworks, and MCP clients.\n
\n\n\n π www.smriti-memcore.com\n
\n\n[](https://pypi.org/project/smriti-memcore/)\n[](https://www.python.org/downloads/)\n[](LICENSE)\n[](https://www.smriti-memcore.com)\n\n\n\n---\n\n## π‘ What is SMRITI?\n\nSMRITI is a high-performance, neuro-inspired long-term memory layer designed to give AI agents persistent, adaptive recall without blocking their real-time execution loop. \n\nInspired by human Dual-Process cognitive theory, SMRITI splits memory operations into:\n1. **System 1 (Immediate Heuristics)**: Decoupled, millisecond-level ingestion of raw interactions into an append-only Episode Buffer.\n2. **System 2 (Async Consolidation)**: Background LLM-driven consolidation that extracts knowledge graphs, resolves contradictions, identifies skills, and decays weak memories.\n\n---\n\n## βοΈ SMRITI vs. Naive RAG & Vector Databases\n\n| Feature | Naive RAG / Vector DBs | SMRITI Memory Engine |\n|---|---|---|\n| **Latency** | Scales linearly with context size; blocks agent loops | **Sub-5ms ingestion** (System 1); System 2 is asynchronous |\n| **Context Window** | Stuffs raw logs, leading to prompt bloat and distraction | **Miller's Law (7 Β± 2 slots)** capacity-bounded Working Memory |\n| **Data Evolution** | Static embeddings; struggles with contradictions/corrections | **Automatic conflict resolution**, abstraction, and temporal decay |\n| **Relationships** | Flat vector search; no concept of entity links | **Semantic Palace Graph** showing structured Room/Topic associations |\n| **Privacy & Sync** | All-or-nothing storage; complex namespace routing | **Private Rooms** and `private=True` tags natively isolating user syncs |\n\n---\n\n## π Key Capabilities\n\n* π§ **Dual-Process Performance**: Zero-blocking real-time loops. Write immediately, analyze when idle.\n* π **Privacy-First (Private Rooms)**: Create local semantic rooms whose memories are automatically excluded from shared/team-wide sync.\n* π **Model Context Protocol (MCP)**: Native MCP server integration with Claude Code, Claude Desktop, Gemini Antigravity, and Codex.\n* π¦ **AMP v1.0 Spec Compliant**: Drop-in compatibility with any agent framework conforming to the Agent Memory Protocol.\n* π **Visual Graph Explorer**: Clean D3.js-based visualization interface with Prometheus metrics monitoring.\n* π **Obsidian Vault Integration**: Automatically syncs your agent's memory graph into an Obsidian vault for human curation.\n* 𧩠**Framework Agnostic**: Integrates natively with LangChain, LlamaIndex, CrewAI, and AutoGen.\n\n---\n\n## π§ Core Architecture\n\n```text\n βββββββββββββββββββββββββββββββββββ\n β Asynchronous Consolidation β\n β (8 Background Processes) β\n β β’ Chunking β’ Cross-Ref. β\n β β’ Conflict Res. β’ Skill Ext. β\n β β’ Forgetting β’ Spaced Rep. β\n β β’ Reflection β’ Defragment. β\n ββββββββββββββββββ¬βββββββββββββββββ\n β background\n ββββββββββββ ββββββββββββ βββββββββββββΌββββββββββ ββββββββββββ\n β Input ββββΆβ Attention ββββΆβ Episode Buffer ββββΆβ Semantic β\n β Text β β Gate β β (append-only log) β β Palace β\n β βββββββββ β (salience β βββββββββββββββββββββββ β Graph β\n β β filter) β β G=(V,E) β\n ββββββββββββ ββββββββββββ ββββββ¬ββββββ\n β\n ββββββββββββ ββββββββββββ βββββββββββββββββββββ β\n β Query ββββΆβ RetrievalββββΆβ Working Memory βββββββββββββ\n β β β Engine β β (7 Β± 2 slots) β\n ββββββββββββ β Q(v) = β βββββββββββββββββββββ\n β Ξ²βcos + β\n β Ξ²βdecay+ β βββββββββββββββββββββ\n β Ξ²βfreq + ββββΆβ Meta-Memory β\n β Ξ²βsal β β (confidence map) β\n ββββββββββββ βββββββββββββββββββββ\n```\n\n---\n\n## π Quick Start\n\n### 1. Unified MCP Server (Claude Code, Gemini, Codex)\n\nSMRITI can be used as a global, persistent memory layer across all your MCP-enabled developer clients.\n\n#### Method A: One-Line Installer (Recommended)\nRun the setup script directly in your terminal:\n```bash\nbash <(curl -s https://raw.githubusercontent.com/smriti-memcore/smriti-memcore/main/install_smriti_mcp.sh)\n```\n\n#### Method B: Via PyPI\nInstall the package and run the setup CLI:\n```bash\npip3 install smriti-memcore\nsmriti_install\n```\n\n---\n\n### 2. Python SDK\n\nFor application developers building custom agent loops.\n\n```bash\npip install smriti-memcore[faiss] # FAISS is recommended for accelerated vector search\n```\n\n```python\nfrom smriti import SMRITI, SmritiConfig\n\n# Initialize memory engine with OpenAI\nconfig = SmritiConfig(\n storage_path=\"./my_agent_memory\",\n llm_model=\"gpt-4o\",\n openai_api_key=\"your-api-key-here\"\n)\nmemory = SMRITI(config=config)\n\n# Ingest observations\nmemory.encode(\"User prefers using PyTorch for neural networks.\")\nmemory.encode(\"User is allergic to shellfish.\", context=\"medical\")\n\n# Recall relevant context using multi-factor retrieval\nresults = memory.recall(\"What framework does the user prefer?\")\nfor mem in results:\n print(f\"[{mem.strength:.2f}] {mem.content}\")\n\n# Manually trigger System 2 background consolidation\nmemory.consolidate()\nmemory.save()\n```\n\n---\n\n## π οΈ MCP Tool Reference\n\nSMRITI exposes **19 tools** (13 native + 6 AMP aliases) for clients:\n\n### Core Tools\n\n| Tool Name | Description |\n|---|---|\n| `smriti_encode` | Ingests a new memory. Accept `private=True` to exclude from team syncs. |\n| `smriti_recall` | Retrieves memories using semantic and graph-based retrieval. |\n| `smriti_get_context` | Helper to inject the current active working memory slots into the context window. |\n| `smriti_how_well_do_i_know` | Performs a meta-memory confidence check on a given topic. |\n| `smriti_knowledge_gaps` | Identifies topics the agent has identified it needs more information on. |\n| `smriti_pin` | Marks a memory as permanent (protects it from strength decay). |\n| `smriti_forget` | Soft-deletes/archives a memory, leaving a cryptographic tombstone. |\n| `smriti_consolidate` | Triggers a background System 2 consolidation run. |\n| `smriti_stats` | Returns system-wide statistics (total memories, rooms, private counts). |\n| `smriti_create_private_room` | Spawns a private room. All memories inside this room are visibility-isolated. |\n| `smriti_open_ui` | Launches the interactive visual D3.js memory graph in your default browser. |\n| `smriti_sync_obsidian` | Exports the Semantic Palace graph structures to markdown files in an Obsidian Vault. |\n\n### AMP v1.0 Alias Tools\nThese endpoints ensure complete conformance with the standard Agent Memory Protocol specification:\n\n| AMP Tool | Native Mapping | Return Format |\n|---|---|---|\n| `amp.encode` | `smriti_encode` | AMP standard JSON response |\n| `amp.recall` | `smriti_recall` | Array of `{id, content, score, timestamp, status}` |\n| `amp.forget` | `smriti_forget` | `{status: \"forgotten\" \\| \"not_found\"}` |\n| `amp.stats` | `smriti_stats` | `{memory_count, ...}` |\n| `amp.pin` | `smriti_pin` | `{status: \"pinned\" \\| \"not_found\"}` |\n| `amp.consolidate` | `smriti_consolidate` | `{status: \"ok\", memories_processed: int}` |\n\n---\n\n## π Framework Integrations\n\n### LangChain Integration\nUse `SmritiLangChainMemory` as a drop-in replacement for default chat buffers. It limits active context using Working Memory and offloads the conversational history to the Semantic Palace graph in the background.\n\n```python\nfrom langchain.chains import ConversationChain\nfrom smriti.integrations.langchain_memory import SmritiLangChainMemory\nfrom smriti import SMRITI\n\nsmriti_engine = SMRITI(storage_path=\"./langchain_smriti_db\")\nsmriti_memory = SmritiLangChainMemory(smriti_client=smriti_engine, top_k=3)\n\nconversation = ConversationChain(\n llm=my_llm,\n memory=smriti_memory,\n)\nconversation.predict(input=\"I prefer backend APIs in Python.\")\n```\n\n---\n\n## π Benchmarks & Performance\n\n### 1. LoCoMo (Multi-System Context Retrieval)\nTested against four architectures on the [LoCoMo](https://github.com/snap-research/locomo) long-context dialogue dataset (28 turns, 15 evaluation questions):\n\n| System | F1 Score | Latency | Tokens/Query | Consolidation |\n|---|---|---|---|---|\n| FullContext | **0.345** | 1147ms | 550 | β |\n| MemGPT-style | 0.334 | 1397ms | 478 | β |\n| NaiveRAG | 0.312 | 1387ms | 145 | β |\n| **SMRITI** | 0.279 | 1317ms | **146** | 41.2s (async) |\n| Mem0-style | 0.235 | 1088ms | 106 | β |\n\n*SMRITI retains high recall while drastically reducing query context size. Consolidation runs in the background and does not block client interactions.*\n\n### 2. LongMemEval (Long-Term Chat Sessions)\nEvaluated over 50+ chat sessions using the [LongMemEval](https://github.com/xiaowu0162/LongMemEval) harness:\n\n| System Configuration | Exact Match Accuracy | Average Query Latency |\n|---|---|---|\n| Baseline (Full Context) | **100.0%** | 11.98s |\n| **SMRITI Dual-Process** | **80.0%** | **0.98s** (12Γ latency reduction) |\n\n---\n\n## βοΈ Configuration Parameters\n\nInitialize `SmritiConfig` with custom parameters to tune the cognitive weights:\n\n```python\nfrom smriti import SmritiConfig\n\nconfig = SmritiConfig(\n working_memory_slots=7, # Capacity limit (Miller's Law)\n \n # Retrieval scoring weights (sum to 1.0)\n recency_weight=0.2,\n relevance_weight=0.4,\n strength_weight=0.2,\n salience_weight=0.2,\n\n # Forgetting & Temporal Decay\n decay_rate=0.99, # Strength multiplier per day\n strength_hard_threshold=0.05, # Memories dropping below this are forgotten\n \n # Palace Graph\n room_merge_threshold=0.85, # Cosine similarity for auto-merging semantic rooms\n)\n```\n\n---\n\n## π Citation\n\nIf you use SMRITI in your research, please cite our technical paper:\n\n```bibtex\n@article{tyagi2025smriti,\n title={SMRITI: A Scalable, Neuro-Inspired Architecture for Long-Term Event Memory in LLM Agents},\n author={Tyagi, Shivam},\n year={2025},\n doi={10.13140/RG.2.2.25477.82407}\n}\n```\n\n---\n\n## π License\n\nSMRITI is licensed under the MIT License. See [LICENSE](LICENSE) for details.\n", "bytes": 10946, "sha": "c15ee469161abe0fd1bb03fcf527ac1b195b3387b97d77fe638b56e0952c3d17", "repo_slug": "smriti-memcore/smriti-memcore", "fonte": "repo", "truncated": false, "api": "https://agentalog.com/api/listings/mcp_io_github_shivamtyagi18_smriti_memory_6e4d4cf2/readme" }