Trajectory learning

Package scope and prerequisites

This catalog ships instructions, not the executable runtime. Python modules, scripts, local services and optional tools named below must be installed separately from the upstream project. Check their availability before running commands; if they are missing, explain the prerequisite instead of claiming execution or results. The host agent may use model tokens while following these instructions.

Overview

Stocke les trajectoires d'optimisation des solveurs déterministes. Décrit la recherche de solutions antérieures similaires. Mesurer la latence, vérifier les prérequis et valider la pertinence avant réutilisation.

Compatible avec ~/.hermes/scripts/trajectory_search.py (Hermes Trajectory Memory).

Files

Usage

from skills.trajectory import capture, search, get_stats

# Capturer une trajectoire
tid = capture(
    solver="bin_pack",
    task="pack 3 database backups into 10GB capacity",
    parameters={"items": [("db1", 4), ("db2", 7), ("db3", 3)], "capacity": 10},
    result={"bins": [...], "bin_count": 2},
    latency=0.002,
    tokens_saved=500,
)

# Rechercher des trajectoires similaires
hits = search("pack items into capacity")
for h in hits:
    print(f"[{h['score']:.2f}] {h['trajectory']['task']}")

# Statistiques
stats = get_stats()
print(f"Total: {stats['total']}, Tokens saved: {stats['total_tokens_saved']}")

Integration with solvers

Dans skills/solvers/solvers.py, chaque fonction peut capturer sa trajectoire:

from skills.trajectory import capture

def assign_balanced(tasks, workers):
    result = ...  # existing logic
    capture("assign_balanced", f"assign {len(tasks)} tasks to {len(workers)} workers",
            {"tasks": tasks, "workers": workers}, result, latency=latency())
    return result

Integration with Hermes

Les fichiers sont compatibles: Botte peut lire les trajectoires Hermes et vice-versa:

# Lire les trajectoires Hermes
from skills.trajectory import load
hermes_trajs = load("~/.hermes/trajectory_store/trajectories.jsonl")

Pitfalls

• Rotation automatique à 5000 entrées
• TF-IDF uniquement — pas de sémantique profonde
• Latence de capture < 1ms (append-only)
• Thread-safe pour usage concurrent