Palace of Truth Project Status

Source basis: Palace of Truth public repository state and merged changes from 6d4e6f8 through 62b4328, reviewed July 9, 2026.

Current Position

Palace of Truth is public at github.com/palaceoftruth/palaceoftruth, open source under AGPL-3.0-only, and positioned as a governed, source-backed truth and memory control plane for agent teams.

The primary agent story is trust before action: scoped startup context, semantic and temporal memory, provenance, stale-source warnings, policy boundaries, and operator review through one self-hostable REST/MCP platform.

See Source-backed wakeup for that workflow and Palace Evidence And Benchmarks for the current public benchmark ceiling.

Shipped Agent Memory And Trust Surfaces

• Durable agent, workspace, session, and tenant-shared memory scopes through REST and MCP.
• Source-backed wakeup with trust states, stale warnings, privacy boundaries, checkpoint pointers, and safe follow-up probes.
• Strict semantic recall with scope isolation, provenance, temporal valid_at filtering, supersession ordering, and bounded output.
• First-class semantic memory entries with temporal and provenance-bearing fields.
• Per-scope retain_mission and quiet_recall profiles plus mission-steered retention.
• Semantic recall exposed through MCP and Hermes, including an optional pre-turn recall path.
• Decision-claim support, stale-source invalidation, operator promotion/rejection/stale/demotion review, and answer audit.
• Wakeup source-trust health in Control Tower.
• Opt-in reflection candidates with source lineage and conflict metadata; generated candidates remain outside canonical recall until review.
• Duplicate replay metadata and durable job/polling contracts for idempotent memory writes.
• Governed multi-agent retrieval with deny-by-default delegated cross-agent policies and audit diagnostics.

Auth And Integration Surfaces

• Repo-owned stdio MCP remains the preferred local Codex and repo-runtime transport.
• OAuth protected-resource discovery, tenant/resource binding, and capability authorization support deployed/shared MCP runtimes.
• Default agent scope configuration reduces repeated setup while preserving server authorization.
• Raw API-key memory clients use explicit read/write scope headers and narrow write grants.
• Packaged Palace agent-client plugin surfaces support Codex/Claude-style installs and Palace-first operating skills.
• Hermes Palace memory plugin 1.0.24 is the current public manifest version.

Broader Product Surface

• Multi-modal ingest for notes, webpages, browser capture, documents, PDFs, images, media, RSS/OPML, source subscriptions, repository sync, and S3-compatible stores.
• Search, chat, feeds, export, graph, jobs, memory APIs, Palace rooms, retrieval diagnostics, and Control Tower operations.
• Worker queues, backpressure telemetry, maintenance sweeps, memory-job recovery, and source-trust summaries.
• Public Helm chart and GHCR packaging for backend, frontend, worker, MCP, bundled Valkey, CloudNativePG, and optional high availability.

Evaluation Boundaries

• Source-backed wakeup and decision-claim governance do not guarantee that every generated answer is fully source-backed.
• Generated reflection is opt-in, review-gated, and generated_unpromoted; it is not an autonomous truth-writing loop.
• Temporal and provenance-aware recall makes context inspectable but does not prove factual correctness.
• Delegated cross-agent reads require explicit server policy; callers cannot grant themselves access.
• New semantic-memory features do not expand the benchmark claims in the evidence page by themselves.
• Authority-sensitive use still needs a defined corpus, validation model, and human review appropriate to the risk.
• Production deletion, cleanup, restore, and admin-secret handling remain operator-led.

What To Evaluate

• Whether wakeup context helps a fresh agent choose a safer first action.
• Whether strict scopes prevent private-agent and unrelated-workspace leakage.
• Whether source and temporal provenance are sufficient for your audit requirements.
• Whether empty, stale, policy-limited, and conflicting cases fail safely.
• Whether recall quality and latency meet your own corpus and runtime budgets.
• Whether local stdio, OAuth-secured MCP, or a managed deployment fits your integration boundary.