What is agenticmemory?

SmartAgentic is a memory layer for AI coding assistants. The README pitches it as a project centric operating system that turns conversations, code context, and technical decisions into a long lived knowledge graph that a team's agents can reason over later. The problem it tries to solve is that current coding assistants forget what was decided in the last session, contradict themselves over time, cannot reuse knowledge across projects, and cannot show why a particular decision was taken. SmartAgentic's answer is a fixed pipeline that runs on every input: extract, rank, resolve context, synthesize, decay or rebase, and sync. The system is organised as several specialised agents. A Micro Agent keeps deep context for a single project. A Meta Agent does global synthesis and consensus across projects. Immune Agents do adversarial review of proposed memory updates. A Sentinel Agent escalates to a human reviewer when something needs sign off. The memory itself is a graph of nodes and edges where the node types are issue, decision, assumption, contradiction, dependency, and abstraction, and edges connect them as parents, related items, contradictions, pointers, or abstraction links. A ranking step assigns a dynamic weight to each node so the agents can pull the relevant subset for a given prompt. The pipeline is split into seven layers visible in the source tree. Layer 0 is an append only event log in JSONL with a hash chain for tamper evidence. Layer 1 runs an LLM extraction step with Vietnamese structured prompts for Qwen, then classifies the node type, scores importance and confidence, pulls tags, and detects parent and related relations. Layer 2 manages a seven state lifecycle with triggers such as human confirmation, usage thresholds, contradiction detection, and time decay, plus a promoter that moves nodes from short memory to long memory. Layers 3 through 6 cover ranking, decay and rebase, context resolution, and synthesis, ending with an agent runtime, a human review loop, and a global sync. Storage is tiered into hot, warm, and cold buckets to control cost. The backend is Python with FastAPI, Uvicorn, APScheduler, and Pydantic v2. LLM and embedding work runs through Transformers, BitsAndBytes, Sentence Transformers, and LiteLLM. The graph and similarity layer uses NetworkX and the datasketch MinHash LSH library. Data lives in SQLite, Parquet via pyarrow, and JSONL. A separate desktop UI built with Electron, React, TypeScript, Vite, Tailwind, and TanStack Query sits in the frontend folder. Setup is the usual clone, create a virtual environment, install requirements.txt, drop a GEMINI_API_KEY in a .env file, and start the API with python main.py serve. The repository also ships a benchmark suite that compares a baseline coding assistant against the SmartAgentic version on compile and test pass rate, golden score, tokens, and latency, with both a built in 30 step trajectory and a custom trajectory file option. The full README is longer than what was shown.