Cuts coding agent token burn up to 72% continuously via contradiction-free memory & verification.
Scored across 15 tools
Most tools have distinguishable purposes, but several clusters overlap: oxyo_consolidate vs oxyo_sleep_abstract both consolidate memory, oxyo_recall vs oxyo_session_pack both retrieve beliefs, and oxyo_verify_python / oxyo_prove_constraints / oxyo_self_play_distill all sit in the verification space. The descriptions do a good job of clarifying boundaries (e.g. LLM-free consolidation vs trace-to-skill distillation), but an agent must read carefully to pick correctly. Additionally, several tools are heavily multi-mode (oxyo_goal, oxyo_causal_infer), which blurs single-purpose expectations.
All tools carry the same oxyo_ prefix and use snake_case, giving a predictable surface. Most follow a verb_noun shape (oxyo_remember_belief, oxyo_skill_register, oxyo_causal_infer), with a few noun-only deviations (oxyo_goal, oxyo_belief_revisions, oxyo_session_pack) that are still readable. Minor inconsistency, not confusing.
15 tools is at the top of the comfortable range but justified by a genuinely broad domain (memory, causality, goals, skills, sandbox verification). Each tool maps to a real capability rather than padding, though a few could be merged (consolidate/sleep_abstract). Slightly heavy but well scoped.
The memory lifecycle is well covered: write (belief, episode, causal link), read (recall, session_pack, belief_revisions), maintain (consolidate, sleep_abstract), plus goals, skills, and verified execution. The main gap is explicit deletion/forgetting or pruning of beliefs/episodes and no dedicated skill-list or memory-stats operation, though supersession and recall partially compensate.
Inspect the immutable audit trail of superseded contradictions vs current truth.
Inputs: limit.
Run causal reasoning across Pearl's Ladder: multi-hop intervene/diagnose, graph-mutilated do_calculus, Rung-3 counterfactual (Abduction-Action-Prediction), or active experiments ranked by Expected Information Gain.
Inputs: do, mode, scope, entity, +2 more.
Record or reinforce a directed cause-effect link in Oxyo's causal graph.
Inputs: cause (required), effect (required), scope, action, +2 more.
Run zero-LLM hippocampal memory consolidation to archive redundant episodes and promote recurring patterns.
Inputs: scope, keep_recent.
Manage persistent multi-session Long-Horizon Goal DAG: create goals, decompose sub-goals, query next unblocked actions, and update status with automatic Causal Backtracking on failures.
Inputs: limit, scope, title, action, +11 more.
Run exact zero-LLM symbolic constraint solving (satisfy + minimize/maximize) or universal invariant theorem proving (prove: verifies assumptions => claim or synthesizes a minimal counterexample).
Inputs: variables (required), claim, sense, action, +3 more.
Search persistent Oxyo memory: surgical episodic snippets, active contradiction-free beliefs (zero-noise filtered), cause-effect chains, and verified Python skills.
Inputs: query (required), scope, k_beliefs, k_episodes, +3 more.
Persist a durable fact, rule, or user preference. Deduplicates paraphrases and supersedes contradicted older 1:1 claims while keeping 1:N additive facts intact.
Inputs: claim (required), scope, confidence.
Store an experience, architectural decision, or observation with dynamic neural salience.
Inputs: content (required), kind, scope, importance, +1 more.
Run autonomous self-play problem verification and curated trace export.
Inputs: seed, difficulty, max_problems, export_corpus, +3 more.
One-call cold-start session capsule (<350 tokens): returns active contradiction-free project/global beliefs, recent actions, causal bug warnings, and verified skills.
Inputs: focus, scope, auto_consolidate.
List or execute a verified Python skill from private or collective Oxyo library.
Inputs: args, name, action.
Register a reusable (and optionally compositional) Python function after verifying 100% of its test cases.
Inputs: code (required), name (required), entrypoint (required), test_cases (required), +3 more.
Consolidate verified execution traces into tested, reusable functions in your persistent skill vault and compact episodic memory.
Inputs: scope, traces, keep_recent, max_new_skills.
Execute pure Python math/algorithmic verification inside Oxyo's isolated sandbox (with optional AST-driven adversarial boundary fuzzing & invariant checks). Assign the final output to RESULT = ...
Inputs: code (required), fuzz, timeout_s, entrypoint, +1 more.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Verified memory for AI agents. Signed assertions, billing attestation, session continuity.
Up to 90% fewer tokens: deterministic, declared tool calls, each with a receipt and an audit trail.
Persistent engineering memory designed to follow developers across coding agents.
Memory that reasons: continual learning for stateful agents. Better context, fewer tokens.
Add one secure layer between your agents and this server.
