Deterministic, self-hosted episodic memory for agents (K3 engine): 1M episodes / 145M tokens per deployment — ~1,450 novels, ~1,130× a
GPT-4o context window, ~27 years of chat. Details: CAPACITY.md. LoCoMo QA 70.49% under equal protocol, reproducible evaluation glue. No implementation sources are included by design. Closed source — see NOTICE.
Try it: an MCP server with a free trial is coming — run infiniteMem from Claude Desktop and your agents without sharing data externally. Contact via repository for trial access.
paper/
— scientific report v2 (LoCoMo + LongMemEval, stronger-reader rows, world-model gate, reranker study, limits, references).harness/
— evaluation scripts only (dataset fetch checks, fresh embedding via NVIDIA API, numpy retrieval crosscheck, QA answer+judge via OpenRouter, rerank gate). They drive the closed engine as a black box and reveal nothing about its internals.results/
— certified numbers:leaderboard.json
(LoCoMo, SOTA2-ready),`leaderboard_lmem.json`
(LongMemEval-S),`readers.json`
(stronger-reader rows + world-model gate), per-question`qa_cache.jsonl`
(3070 records, no gold texts), retrieval and QA reports, reranker negative-result report.
- Fetch
snap-research/locomo
data/locomo10.json
(CC BY-NC 4.0, non-commercial, cite Maharana et al., ACL 2024) and verifySHA256 79FA87E9…
. - Set env keys (never commit):
NVIDIA_API_KEY
(embeddings`llama-nemotron-embed-vl-1b-v2`
, 768-D, passage`Speaker: text`
/ query raw) andOPENROUTER_API_KEY
(answer+judge`openai/gpt-4o-mini-2024-07-18`
, temp 0, verbatim Mem0 prompts). - Run harness in order:
exnovo_embed.py
→eval_fresh_retrieval.py
→`exnovo_qa.py --mode k3`
+`--mode oracle`
. Expected: retrieval78.85/74.10
, QAk3 70.49 / oracle 84.89
(±0.5pp API variance). Black-box engine, brute top-10 single-shot, dedup off, no rerank, 1 run. Answer/judge models and prompts pinned (see paper §3, §6). Dataset CC BY-NC 4.0. Contact via repository for engine access.