OpenLunum
OpenLunum develops Lunum, an intermediate representation for agent memory and context. For single sentences, an agent proposes a structured meaning and core code decides whether it earns a versioned identity (lfp:2.1), abstaining rather than guessing. For long text, deterministic discourse records split messages and tool output into source-span units. The source text is always kept. Pre-1.0 research (contract 0.19), not a production dependency.
- Independent fresh-sentence evaluation (round 2): 27/30 on outcome, 20/30 on meaning vs 7 parses dropped thresholds or dates, or used a wrong verb, and still got an identity (evidence)
- Discourse context, offline replay of 15 real sessions: mean prompt tokens 5,860 vs 19,586 without it (−70%; −55% without the one point over the limit; 13 of 52 points got larger) (evidence)
- Exact recall with Greek diacritic folding (recall@5): 0.992 for unaccented and inflected Greek vs 0.283 before folding; the 170 control items unchanged (evidence)
Source on GitHub