trellis-ai

name	trellis-ai
description	Structured institutional memory for AI agents — traces, precedents, knowledge graph, temporal versioning, and governed mutations. Complements built-in memory with structured knowledge.
version	1.0.0
status	preview
homepage	https://github.com/ronsse/trellis-ai
metadata	{"openclaw":{"requires":{"anyBins":["trellis-mcp"]}}}

Trellis

Status: preview. Tool signatures and prompts are in flux while parallel work lands. Expect revisions before the next minor release.

Structured institutional memory for AI agents. Record traces of your work, build a shared knowledge graph, and retrieve context packs before starting new tasks.

When to Use Trellis vs Built-in Memory

Use Case	Built-in Memory	Trellis
Daily notes, session context	Yes
Quick personal reminders	Yes
Structured traces of work (steps, tool calls, outcomes)		Yes
Reusable patterns and precedents		Yes
Knowledge graph with typed relationships		Yes
Temporal versioning (time-travel queries)		Yes
Cross-agent institutional knowledge		Yes
Evidence linking (docs, snippets, files)		Yes

Rule of thumb: If it's a note for yourself, use built-in memory. If it's structured knowledge that other agents or future sessions should learn from, use Trellis.

Available Tools

11 macro tools, all returning token-budgeted markdown (not raw JSON). Eight cover the most common write-and-read workflow; three more provide sectioned context for richer multi-step retrieval.

Core tools

Tool	What It Does	Example
get_context	Search docs + graph + traces for task context	get_context(intent="implement retry logic", domain="backend")
save_experience	Record a trace of completed work	save_experience(trace_json="{...}")
save_knowledge	Create an entity in the knowledge graph	save_knowledge(name="auth-service", entity_type="service")
save_memory	Store a document for later retrieval	save_memory(content="Rate limiting uses token bucket algorithm")
get_lessons	List precedents (proven patterns)	get_lessons(domain="backend", max_tokens=1500)
get_graph	Explore entity neighborhood	get_graph(entity_id="01JRK5N7QF", depth=2)
record_feedback	Record whether a task succeeded	record_feedback(trace_id="01JRK5N7QF", success=true)
search	Search documents and entities	search(query="database migration", limit=5)

Sectioned-context tools (use for richer multi-step or workflow-spanning retrieval)

Tool	What It Does	Example
get_objective_context	One pack covering domain knowledge + operational context for a whole workflow	get_objective_context(intent="ship auth migration", domain="backend")
get_task_context	Pack scoped to specific entities for one step inside a workflow	get_task_context(intent="rotate JWT keys", entity_ids=["auth-service"])
get_sectioned_context	Pack with caller-defined sections, per-section budgets and affinities	get_sectioned_context(intent="...", sections=[{...}, {...}])

Patterns

Retrieve Before Acting

Before starting non-trivial work, check for prior art:

get_context(intent="what you're about to do", domain="relevant-domain")

This returns relevant traces, precedents, and evidence so you avoid repeating past mistakes and reuse proven patterns.

Record After Success

After completing meaningful work, save the experience:

save_experience(trace_json='{"source": "agent", "intent": "what you did", "steps": [...], "outcome": {"status": "success", "summary": "what happened"}, "context": {"domain": "backend"}}')

Then record feedback:

record_feedback(trace_id="<returned_id>", success=true, notes="Clean implementation")

Link Evidence

When you discover useful documentation or patterns, store them and connect to entities:

save_memory(content="API rate limits: 1000 req/min per client", metadata={"source": "api-docs", "domain": "platform"})
save_knowledge(name="rate-limiter", entity_type="concept", relates_to="api-gateway-id", edge_kind="entity_part_of")

Context Window Tips

• Use max_tokens to control response size — default is 2000 tokens
• For quick lookups, set max_tokens=500
• For deep research, allow up to max_tokens=4000
• All responses are pre-formatted markdown, ready for your context window
• Prefer get_context for broad task context; use search for targeted queries