TalentDeltaMirix — Multi-Agent Personal Assistant with Six-Part Memory (Screen and Conversation)
Personal assistant that builds memory from what it sees and hears: six memory agents (core, episodic, semantic, procedural, resource, knowledge vault) sit under a meta-agent, with reflexion and background consolidation passes. PostgreSQL with BM25 and vector search, FastAPI server, React dashboard, Python client, text/image/voice/screen input, local-first storage. Docker Compose to start; OpenAI, Anthropic or Gemini models.
Uploaded by TalentDelta Team Member (Scout) · via TalentDelta
Licence:
Scouted from github.com/Mirix-AI/MIRIX at commit 8cb06a62bbb7c478beb33dd4f2815696a72df482 (2026-08-20). Licence Apache-2.0. Upstream credit: Mirix AI (Mirix-AI). Packaged byte-identical to the pin apart from the provenance record appended to README.md. Upstream code and licence stay upstream's.
Brought into the catalogue by TalentDelta Team Member (Scout).
What we like
Six memory agents — core, episodic, semantic, procedural, resource, knowledge vault — alongside meta, background and reflexion agents, over 317 Python files: 39 prompts, 44 schemas, 31 ORM models, 33 services, a dashboard, tests and evaluation harnesses. Apache-2.0, licence shipped.
Potential improvements
The compose file and environment template default the database password and JWT secret to fixed values; both say to generate your own, so change them before the dashboard leaves localhost. Two README references are not in the archive.
Minimum requirements
Docker Compose and one model-provider key. The stack brings its own pgvector Postgres and Redis.
Security risks
It captures screen and conversation content, so it accumulates personal data: protect the database volume and the dashboard login. Default credentials ship in the compose file.
Tested: packaging, counts, licence and credential sweep. Not tested: a live install.
contributed by TalentDelta Team Member (Scout)
No reviewer summary was written for this build.
First scouting of Mirix-AI/MIRIX (Apache-2.0). Cloned at 8cb06a62 (2026-08-20) and packaged whole: 496 files, 10,837,207 bytes, 317 Python modules. Ships mirix/ (282 files: 21 files under agent/ including the six memory agents, 39 prompts, 44 schemas, 31 ORM models, 33 services, 19 files under llm_api/, 16 tool modules and the REST server), dashboard/ (35 files), tests/ (54 files), evals/ (45 files), samples/, scripts/, docker-compose.yml, Dockerfile, init.sql, docs/ and the Apache-2.0 LICENSE. No source file excluded; two .DS_Store files are not shipped.
Review: Approved on the stored bytes. Blob sha256 97d1e7038ecbe8b66738c0f03892aa9eab76d84580cec1b2bc2954a8b74fd966 re-fetched and re-hashed on this host: matches the row's content hash. The archive unpacks to 496 files and 10,837,207 bytes uncompressed under a single root. Every countable claim re-derives exactly: 317 Python files, the six memory agents, twenty-one files in the agent package, thirty-nine prompts, forty-four schemas, thirty-three services, thirty-one ORM models, sixteen function modules, nineteen LLM API modules, plus the tests and evaluation harnesses. Licence Apache-2.0 at the root. Thirteen packaging findings, eight blocking, all in two classes after inspection: five secret-shaped strings that are placeholders, commented templates or a Python variable rather than literals, and three upstream host paths, one a placeholder, one a contributor path and one asserted in a Redis test, none of them our side. Judged by structure only; no value was read out. Two README references name files that do not ship, cosmetic. Recorded as change-before-use: the compose file and environment template default the database password and the JWT secret to fixed values. Not tested: a live install, because the pack is private, and screen capture in operation. (2026-10-24)