- Letta agents use memfs memory (git-backed context files) and have no per-agent fallback model or bundled Ollama — those are Hermes-only. + Letta agents keep their memory in core-memory blocks on the shared server. Letta's git-backed memory files (memfs) are not enabled on this deploy path yet, so the context-files view will be empty. Letta agents also have no per-agent fallback model and no bundled Ollama — those are Hermes-only.
A Letta agent is a stateful brain on the shared server — it's reached over REST, not Signal/Telegram/Slack directly. Multiplayer messaging arrives via a separate Hermes front that proxies a group conversation to this agent (v1 bridge). Nothing to configure here.
++ Through that front, group approval and the audit trail carry over — they gate the turn before it runs. Budget enforcement does not carry over at all. The spend check sums the token columns an agent writes to its own state DB, and the Letta bridge never writes them, so proxied turns count as zero spend on every path. Do not rely on the front's budget cap to bound a Letta agent. +
)} diff --git a/components/harness/letta-agent-detail.tsx b/components/harness/letta-agent-detail.tsx index 1ef943c..df4c638 100644 --- a/components/harness/letta-agent-detail.tsx +++ b/components/harness/letta-agent-detail.tsx @@ -11,14 +11,20 @@ * Lifecycle buttons → Send message (Reset/Delete deferred to a later phase) * Overview CPU/mem → model handle + agent identity * Logs tab → message / turn viewer (send + rendered reply) - * Memory rows → core-memory blocks + memfs context-file view (A3) + * Memory rows → core-memory blocks (+ a context-file view, if enabled) * - * MEMFS (2026-07-21 research, memory/specs/2026-07-21-letta-memfs-api-and-a3.md): - * a modern Letta agent's memory is a git-backed file tree; `system/` files are - * pinned into the prompt, the rest open on demand. We surface BOTH the - * core-memory blocks (`/core-memory/blocks`) and the live file view (`/files`). - * Everything here is read-only except "send a message". Block/file WRITES and - * the memfs↔REST live-sync semantics need a live server (Phase 4). + * MEMFS IS NOT ENABLED — DO NOT WRITE COPY THAT PROMISES IT. + * The 2026-07-21 research (memory/specs/2026-07-21-letta-memfs-api-and-a3.md) + * describes memfs as a git-backed memory file tree with `system/` files pinned + * into the prompt. That is Letta's capability, not ours yet: we never send + * `git_enabled` (zero occurrences in executable code — only doc comments like + * this one), so agents created by our deploy path take the server default, + * observed as `git_enabled: false`. So the `/files` view below is + * normally EMPTY, and its empty state must say "not enabled" rather than imply + * the feature is there and the agent simply has no files. + * + * Everything here is read-only except "send a message". Block/file WRITES, + * enabling memfs, and the memfs↔REST live-sync semantics are all unshipped. */ import { useState } from 'react' @@ -77,8 +83,8 @@ export function LettaAgentDetail({ harness }: { harness: Harness }) {{isServer - ? 'Letta server (a real container). Hosts N agents as Postgres rows — start/stop/logs via Docker are available on the server, agents are managed over REST.' - : 'Letta agent — a row in the Letta server’s Postgres, not a container. No CPU/memory, no lifecycle buttons; managed over REST.'} + ? 'Letta server (a real container). Hosts N agents as Postgres rows — start/stop/logs via Docker are available on the server; the agents themselves are reached over REST.' + : 'Letta agent — a row in the Letta server’s Postgres, not a container. No CPU/memory, no lifecycle buttons. Read-only over REST, plus sending a message: editing memory, reconfiguring and deleting are not shipped.'}
{/* Overview: model handle + identity, not CPU/mem (design §4a) */} @@ -131,14 +137,18 @@ function ServerAgentsList() { ) } -/** Core-memory block viewer — the seed layer that projects into memfs system/ files (design §4a). */ +/** + * Core-memory block viewer. This is the agent's actual live memory surface on + * our deploy path, because memfs is not enabled — the "blocks project into + * system/ files" story only applies once git-backed memory is turned on. + */ function MemoryBlocks({ agentId }: { agentId: string }) { const { data: blocks, loading, error } = useApi
- Read-only. Under memfs these blocks project into system/ files (e.g. persona.md). Editing lands with the live server (Phase 4).
+ Read-only — editing memory blocks from Swarm Map is not shipped. These blocks are this agent's memory: git-backed memory files (memfs) are not enabled, so nothing is projected into system/ files yet.
Loading…
} {error &&Couldn't read memory: {error}
} @@ -170,25 +180,44 @@ function MemoryBlocks({ agentId }: { agentId: string }) { } /** - * memfs context-file view — the "what's in this agent's memory right now" - * surface (GET /v1/agents/{id}/files). `system/` files are pinned into the - * prompt every turn; others open on demand. The open/pinned marker is a - * CONTEXT-LOADING indicator, NOT an access-control boundary (memfs has no - * per-file ACL — see the airlock note in the spec). + * Context-file view (GET /v1/agents/{id}/files) — Letta's memfs surface. + * + * HONESTY: memfs is NOT enabled on our deploy path (`git_enabled: false`), so + * this list is empty for every agent we create. The panel therefore leads with + * an explicit "not enabled" state instead of an empty box under a heading that + * promises git-backed memory. The file list still renders if a server does + * expose one, so this stays useful for externally-created agents. + * + * When files ARE present: `system/` files are pinned into the prompt every + * turn; others open on demand. The open/pinned marker is a CONTEXT-LOADING + * indicator, NOT an access-control boundary (memfs has no per-file ACL — see + * the airlock note in the spec). */ function ContextFiles({ agentId }: { agentId: string }) { const { data: files, loading, error } = useApi
- The agent's git-backed memory file tree. Pinned = a system/ file loaded into the prompt every turn; open = currently in context. This is a loading indicator, not a permission boundary.
-
+ The agent's git-backed memory file tree. Pinned = a system/ file loaded into the prompt every turn; open = currently in context. This is a loading indicator, not a permission boundary.
+
Loading…
} {error &&Couldn't read context files: {error}
} {files && files.length === 0 && ( -No context files (agent may predate memfs, or files aren't exposed on this server).
++ Git-backed memory files are not enabled for agents Swarm Map deploys, so there is nothing to show here. Core memory (blocks) above is this agent's live memory. Enabling memfs is not shipped yet. +
)} {files && files.length > 0 && (